⏳ This skill is pending AI review.

Scores will appear once the review pipeline completes.

v1.0.0

image

smixs@smixs⭐ 667 stars· last commit 9d ago· 1 open issues

>

Choose how to use this skill

You do not need every option. Choose the path your AI client supports. The stable page stays the same; versioned files are immutable.

1. Native installer

This listing has no registered native installer command. Use the complete package or source fallback below, depending on what your client supports.

Do not guess an installer command or replace an existing version without reviewing the diff.

2. Complete package recommended

Download the ZIP when available. It includes SKILL.md plus the references, security notes and version metadata.

No complete ProSkills package is published for this listing yet.

3. Prompt-only

Copy the prompt above when the agent can read the stable page or when you want to adopt the workflow without installing a skill.

Need only the instruction file?

Download SKILL.md only if your client requires a single file. The complete ZIP is safer for a full installation because it preserves the references and release context.

No path installs or executes anything by itself. Your agent still needs access to the project files. Before updating, compare the installed version and review the diff.

—/10

// RATINGS

⭐GitHub Stars
⭐⭐⭐⭐ 667 on GitHubGitHub ↗

Popular

🟢ProSkills Score
—
📍

Not yet listed on ClawHub or SkillsMP

// README

🎬 Visual Skills — AI Film Director for Your Movie

Visual Skills — AI film director for your agent: a directors chair, a camera on a dolly and a figure with an umbrella on a rainy neon street

skills.sh Claude Skill License: CC BY 4.0

Two Claude Skills that turn your agent into a working film crew: video writes AI video prompts the way a director, screenwriter and editor would; image writes image prompts the way an art director would. Both pick the right model for the task, apply its exact syntax, and return a copy-paste-ready prompt.

Most prompting guides teach you syntax. This one teaches your agent cinema — and that is what makes it the strongest tool available for directing AI video.

Dramaturgy first, syntax second

[!IMPORTANT] Model syntax is worth nothing until the dramaturgy is there. Editing, staging, camera, light, the objects allowed in frame — hard rules, all of them written into the skill. That is what makes it a director instead of an autocomplete for adjectives. Get them right and the model finally has something worth rendering; get them wrong and no amount of correct syntax saves the shot.

The heart of the video skill is video/references/dramaturgy.md — how films are actually built, compressed into rules an agent can execute on a 5-30 second clip. Same idea, both columns below. Only one of them can be filmed.

cinematic shot of a man
in a kitchen at night,
epic lighting, moody
atmosphere, 4k
Emotion: hunger as loneliness. Object: last sausage.
Final image: fridge light dying on his face.

Shot 1 · 0.0-1.6s · wide, 24mm, static
Dark kitchen. He stands with one hand on the fridge
door, not opening it. Only the wall clock moves.

Cut · 1.6-3.4s · medium, 50mm, push-in
The push starts the frame he sees the shelf is empty
— that is what changed. Cold blue light, jaw sets,
one stomach growl, then nothing.

Cut · 3.4-5.0s · macro, 100mm
Two fingers close on the last sausage. Half-beat hold.
Door swings shut, the light dies on his face.

No desire, no obstacle, no geometry, no cut, no final image. The model picks all five for you — and picks differently on every run.

Every line is a physical fact a camera could record: a reason for the move, a body carrying the emotion, a sound, an object, an ending. Nothing left for the model to invent.

[!CAUTION] Banned everywhere: cinematic · epic · stunning · masterpiece · beautiful lighting · dynamic camera · he is sad. Each one is a placeholder for a detail the writer failed to invent, and not one of them renders.

0 1  ·  L A WThe scene formula

desire + obstacle + geometry + gaze + rhythm

Five elements. Name each one in a single sentence before a word of prompt is written: what the hero wants right now, what blocks it, who stands where, where the eye is forced to look, how long each shot lives. Anything less is decoration.

0 2  ·  D E T A I LThe Details Law

Every shot owns three physical facts: one environmental pressure (cold refrigerator light, wet asphalt), one micro-action of the body (jaw locks, knuckles whiten), one sound anchor or visual motif.

"He is sad" does not render. A jaw does.

0 3  ·  E D I T I N GWalter Murch's Rule of Six — where to cut, in priority order. Each item outweighs everything below it combined.

Cutting "for pace" is item three. Serving item three ahead of emotion and story is exactly how TikTok mush gets made — and it is the default behaviour of every model you will ever prompt.

0 4  ·  S E L E C T I O NThe three-jobs rule

A shot either changes emotion, advances action, or increases pressure. A shot that does none is deleted, however pretty it came out.

"Beautiful establishing shot" is not a job.

0 5  ·  S T A G I N GBlocking, camera, environment

Fincher — every camera move answers "what changed?", otherwise the camera is static. Spielberg — even in chaos the viewer knows where the hero, the threat and the exit are. Kurosawa — one weather, one pressure, carrying the whole scene.

0 6  ·  R H Y T H MMontage is a staircase

long → shorter → shorter → pause → impact

The pause before the hit matters more than the speed of the cuts. Beat maps for 15 / 30 / 60 / 90 seconds — Hook, Pressure, Crack, Impact, Aftermath. Never skip the Crack.

0 7  ·  S P E CThe shot card and the five anchors

Fourteen fields per storyboard row: framing, composition, camera, movement reason, eye-trace, duration, cut type, sound, light. An empty field is missing direction. Per piece, exactly five anchors — one emotion, one motif, one object, one break, one final image.

None of this is advice the agent is free to skip. dramaturgy.md loads before any model file, and the output is gated twice on the way out — the six-point dramaturgy check and a three-detail audit on every shot. A prompt that fails either one is not returned.

Supported models

// HOW IT'S BUILT

// FILES

25 files shipped with this skill. Inline text is the catalog copy. The upstream link is the source repository.

  • image/SKILL.md8.5 KB21f54f320d56upstream ↗
    ---
    name: image
    license: CC-BY-4.0 (attribution required — Serge Shima, github.com/smixs/visual-skills)
    description: >
      Image prompting skill for Nano Banana (NBP/NB2) and GPT Image 2.5 (Flare/Sunburst). Writes ready-to-use
      prompts with model/quality/size recommendations. Use when: "нарисуй", "сгенерируй
      картинку", "image prompt", "промпт для картинки", blog covers, slides, posters,
      product shots, UI mockups, storyboards, character sheets, edit/colorize, style transfer,
      vision analysis, image-to-prompt, nb, NBP, NB2, gpt-image-2.5, multi-panel grids,
      ecommerce product photography, fashion editorial, food/beverage ads, cinematic portraits.
      Do NOT use for: video (use video skill), 3D models, audio, non-image tasks.
    ---
    
    # Image Prompting — Nano Banana & GPT Image 2.5
    
    This skill writes image prompts. It does not generate images. The output is: model name + quality / size / aspect ratio + the prompt itself.
    
    The body of this SKILL.md is intentionally thin so you cannot fake a result by reading it alone. The actual rules — what the models reward, what they punish, how to phrase a 5-slot template, when to add `quality: high`, when to use image grounding — live only in the reference files.
    
    ## Route first — is this actually an image-prompt task?
    
    - **Motion, clips, montage** (Seedance, Kling, Veo, any image-to-video): use the sibling `video` skill. This skill's storyboard and keyframe outputs feed it.
    - **No idea or script yet** (user wants a concept or an ad scenario, not a picture): if the `creative-director` skill is installed, start there — it develops ideas and scripts for commercials and beyond ([github.com/smixs/creative-director-skill](https://github.com/smixs/creative-director-skill)).
    - **A concrete image is needed** — this skill. Continue below.
    
    ---
    
    # Mandatory reading order — DO NOT WRITE A PROMPT WITHOUT THIS
    
    Past attempts to write prompts directly from this skill body produced lazy, generic results. Each model has its own physics; common rules collapse into mush when applied without model-specific syntax. Read in this order before producing any prompt:
    
    ### Step 1 — always read first → [models.md](references/models.md)
    
    Decide: Nano Banana (NB2 or NBP) or GPT Image 2.5 (Flare for speed, Sunburst for precision edits). The choice changes the prompt syntax fundamentally — natural-language paragraphs vs. labeled 5-slot template, quality settings, which features exist (image grounding only on NB, EXACT TEXT discipline only on GPT Image, etc.).
    
    If the user named a model — confirm and proceed. If not — pick using the table in `models.md`, then state your choice in the output header.
    
    ### Step 2 — read **one** model file (the one you picked)
    
    - **Nano Banana** → [nano-banana.md](references/nano-banana.md)
      Image grounding for real locations. Extreme aspect ratios (1:8, 8:1, 4:1). Thinking mode. JSON for 5+ elements. Up to 14 reference images. Why you must NOT write `50mm / f-stop / ISO` numbers.
    
    - **GPT Image 2.5** → [gpt-image.md](references/gpt-image.md)
      5-slot template (Scene / Subject / Important Details / Use Case / Constraints). Anti-slop banned-words list. `quality: low / medium / high / xhigh / max` as a deliberate fidelity lever. Size constraints (multiples of 16, max 3:1, up to 4K 3840×2160). Two-column edit logic (Change / Preserve / Constraints). Up to 16 reference images with explicit roles.
    
    The model file is non-negotiable. Skipping it is the single biggest cause of weak prompts.
    
    ### Step 3 — always read after the model file → [golden-rules.md](references/golden-rules.md)
    
    Universal rules that apply to both models: start with a verb, positive framing, hex colors, quote text, edit don't re-roll, one change per iteration, reference images.
    
    ### Step 4 — task-shaped reading (load only what matches the request)
    
    Pick zero or more, depending on what the user asked for:
    
    - Text in image, infographic, diagram, multilingual rendering → [text-rendering.md](references/text-rendering.md)
    - Edit existing image (object removal, lighting swap, colorization, restoration, localization) → [editing.md](references/editing.md)
    - Character continuity across multiple images / panels → [characters.md](references/characters.md)
    - **The image must pass as a real photograph** (portrait, reportage, UGC, casting, product-in-hand) — or the user says the result "looks AI", "too glossy", "not like the reference" → [de-slop.md](references/de-slop.md). Model default priors, banned booster words, capture pipeline instead of adjectives, located imperfections.
    - Presentation slides → [slides.md](references/slides.md)
    - Sequential narrative (storyboard, comic, panel sequence) → [storyboards.md](references/storyboards.md)
    - Sketch → final, wireframes, structural input → [structural.md](references/structural.md)
    - 2D → 3D, floor plans, isometric → [dimensional.md](references/dimensional.md)
    - **Vision analysis / image-to-prompt / style transfer from a reference image** → [vision-decomposer.md](references/vision-decomposer.md). Load this whenever the user attaches an image and asks to recreate, match, decompose, or transfer its style.
    - **Multi-panel compositions** (grids, collages, storyboard sheets in ONE image) → [multi-panel.md](references/multi-panel.md). 9-cell TVC grids, 2x2 portrait grids, 3-panel campaign collages, 4x3 borderless grids, 6-frame cinematic sequences, before/after splits, 12-panel storyboard posters.
    - **Industry pattern libraries** — proven prompt templates by vertical. Load the matching file:
      - E-commerce product shots → [patterns/ecommerce.md](references/patterns/ecommerce.md)
      - Fashion editorial campaigns → [patterns/fashion-editorial.md](references/patterns/fashion-editorial.md)
      - Food & beverage advertising → [patterns/food-beverage.md](references/patterns/food-beverage.md)
      - Cinematic portraits → [patterns/portrait-cinema.md](references/patterns/portrait-cinema.md)
      - Posters & illustration → [patterns/poster-illustration.md](references/patterns/poster-illustration.md). Also holds the Style DNA + Reject Checklist method — how to fix a strong visual style in four lines before generating and how to judge the returned image against it. Usable with any pattern in this list, not only posters.
      - Character design (turnarounds, expression sheets, outfit grids) → [patterns/character-design.md](references/patterns/character-design.md)
      - UI mockups & social media formats → [patterns/ui-social.md](references/patterns/ui-social.md)
    
    ### Step 5 — read for production language → [creative-direction.md](references/creative-direction.md)
    
    Studio-quality vocabulary for lighting design, camera and hardware, color grading and film stock, materiality and texture. Read when you need precise terms beyond what `golden-rules.md` covers.
    
    ### Step 6 — read if structuring a complex prompt → [prompt-framework.md](references/prompt-framework.md)
    
    Universal element checklist (subject, context, action, environment, camera, lighting, mood, materials, palette, format), detail modes (concise / standard / verbose / cinematic verbose), parameterized templates, output structure with parameters and exclusions.
    
    ---
    
    # Output format
    
    When you return the prompt, structure it like this:
    
    ```
    Model: <nano-banana-2 | nano-banana-pro | gpt-image-2.5-flare | gpt-image-2.5-sunburst>
    Quality: <low | medium | high | xhigh | max>   (only for gpt-image-2.5)
    Size / Ratio: <e.g. 1536×1024 or 16:9>
    
    Prompt:
    <the prompt text, ready to copy>
    
    Notes:
    - <anything you inferred or assumed because the user did not specify>
    ```
    
    For edits, also include an explicit preserve-list (mandatory for gpt-image-2.5, recommended for nano-banana):
    
    ```
    Change: <one concrete thing>
    Preserve: <face, pose, lighting, framing, geometry, ...>
    Constraints: <no extra objects, no drift, ...>
    ```
    
    ---
    
    # Final response style
    
    Prefer: ready-to-copy prompts, hex colors, concrete materials, named compositions, model-specific syntax (5-slot for GPT Image, natural prose for Nano Banana).
    
    Avoid: tag soup ("cool, modern, 4k"), vague praise ("stunning, epic, masterpiece" — actively hurts GPT Image 2.5), negative framing ("no people, no cars" — invert to positive), external comparisons ("like Apple ad" — describe the visual properties instead), numerical lens parameters in Nano Banana prompts (it ignores them).
    
    ---
    
    *Author: Serge Shima ([t.me/aimastersme](https://t.me/aimastersme) · [sergeshima.com](https://sergeshima.com) · [aimasters.me](https://aimasters.me)) · License: CC BY 4.0 — attribution required · Source: [smixs/visual-skills](https://github.com/smixs/visual-skills)*
    
  • image/references/characters.md2.4 KB5069a4a9b4f6upstream ↗
    # Character Consistency
    
    Up to 14 reference images (6 high-fidelity). Identity locking across scenes.
    
    > If the character must read as a real person and the model keeps beautifying them — idealized eyes, poreless skin, model stance — identity locking alone will not hold. Pair this file with [de-slop.md](de-slop.md).
    
    ## Identity Locking
    
    ```
    Using person from [IMAGE REFERENCE]:
    Keep facial features exactly the same.
    Change: [EXPRESSION / POSE / CLOTHING / SETTING]
    Maintain: identity, recognizable features
    ```
    
    Key phrase: "Keep facial features exactly the same as Image 1"
    
    ## Viral Thumbnails
    
    ```
    Design viral thumbnail using person from Image 1.
    
    Face: Keep features same, change expression to [excited/surprised/shocked]
    Action: Pose on [left/right], pointing toward [OBJECT]
    Subject: [OBJECT] on opposite side
    Graphics: Bold [COLOR] arrow connecting person to object
    Text: "[HEADLINE]" in massive pop-style, white outline, drop shadow
    Background: Blurred [SETTING]
    
    High saturation and contrast.
    ```
    
    ## Multi-Character Consistency
    
    ```
    Create [NUMBER]-part story with these [N] characters.
    
    Identity: Keep attire and features consistent throughout.
    Variation: Different angles, expressions, distances per image.
    Rule: Only one of each character per image.
    
    Story arc: [DESCRIPTION with emotional highs/lows]
    ```
    
    ## Brand Asset Generation
    
    ```
    Using this [PRODUCT] as reference:
    Create [NUMBER] [TYPE] shots.
    
    Maintain: Brand consistency, quality level
    Vary: Angles, lighting, context
    Style: [editorial / lifestyle / product]
    
    Generate one at a time.
    ```
    
    ## Expression Control
    
    While maintaining identity:
    - "Change expression to excited and surprised"
    - "Make them look thoughtful, gazing left"
    - "Add confident smile, direct eye contact"
    
    ## Pose/Action Control
    
    - "Pose pointing finger toward [DIRECTION]"
    - "Standing with arms crossed"
    - "Mid-stride, dynamic motion"
    - "Seated, leaning forward"
    
    ## Attire Changes
    
    - "Same person, but wearing [NEW OUTFIT]"
    - "Keep face, change to formal business attire"
    - "Maintain identity, add [ACCESSORY]"
    
    ## Environment Swaps
    
    - "Same person in [NEW SETTING]"
    - "Transport to [LOCATION], keep identity"
    - "Change background to [SCENE], maintain lighting on face"
    
    ---
    
    *Author: Serge Shima ([t.me/aimastersme](https://t.me/aimastersme) · [sergeshima.com](https://sergeshima.com) · [aimasters.me](https://aimasters.me)) · License: CC BY 4.0 — attribution required · Source: [smixs/visual-skills](https://github.com/smixs/visual-skills)*
    
  • image/references/creative-direction.md4.8 KB68b6ac2b3933upstream ↗
    # Creative Direction
    
    Studio-quality controls for lighting, camera, color, and materiality.
    
    ## Lighting Design
    
    Tell the model exactly how the scene is illuminated.
    
    **Studio setups:**
    - "Three-point softbox setup" — even product/portrait lighting
    - "Single key light from upper left" — dramatic shadow
    - "Ring light, frontal" — beauty/fashion, flat lighting
    
    **Dramatic effects:**
    - "Chiaroscuro lighting with harsh, high contrast"
    - "Golden hour backlighting creating long shadows"
    - "Rim lighting, subject silhouetted against bright background"
    - "Volumetric god rays through window"
    - "Neon underglow, cyberpunk street"
    
    **Natural:**
    - "Soft overcast daylight, even diffusion"
    - "Dappled sunlight through trees"
    - "Blue hour, cold ambient light"
    
    ## Camera & Hardware
    
    Specify camera type to change the visual DNA of the image.
    
    **Hardware names that work:**
    - GoPro — immersive, wide-angle distortion, action feel
    - Fujifilm — authentic color science, warm tones
    - Disposable camera — raw, nostalgic flash aesthetic, grain
    - Hasselblad — medium format, shallow DOF, fashion/portrait
    - iPhone — casual, smartphone aesthetic
    - Polaroid — instant film, white border, faded colors
    
    **Angles:**
    - "Low angle" — power, drama
    - "Aerial view" / "bird's eye" — scale, pattern
    - "Eye-level" — neutral, relatable
    - "Dutch angle" — tension, unease
    - "Worm's eye view" — extreme drama
    
    **Shot types:**
    - Close-up, medium shot, wide shot, extreme close-up
    - "Waist-up portrait", "full body", "over-the-shoulder"
    
    **Focus:**
    - "Shallow depth of field" — blurred background
    - "Deep focus" — everything sharp
    - "Rack focus" — shift attention
    - "Tilt-shift" — miniature effect
    
    > Note: Named lens parameters (50mm, f/2.8) are less reliable than
    > descriptive terms. Prefer "shallow depth of field" over "f/1.8".
    
    ## Color Grading & Film Stock
    
    Set emotional tone through color and texture.
    
    **Film stocks:**
    - "Shot on 1980s color film, slightly grainy" — nostalgic
    - "Kodak Portra 400" — warm skin tones, pastel highlights
    - "Fuji Velvia" — hyper-saturated, vivid
    - "Ilford HP5" — classic B&W grain
    - "Cinestill 800T" — tungsten, halation around lights
    
    **Digital grades:**
    - "Cinematic color grading with muted teal tones"
    - "High contrast, desaturated" — gritty, editorial
    - "Warm analog tones, lifted blacks" — vintage
    - "Cool blue shadows, warm highlights" — modern cinema
    - "Cross-processed, shifted greens" — experimental
    
    ## Materiality & Texture
    
    Define physical makeup of subjects. Don't just say the thing — describe what it's MADE of.
    
    | Vague | Specific |
    |-------|----------|
    | "a suit" | "navy blue tweed suit, visible weave" |
    | "armor" | "ornate elven plate armor, etched with silver leaf patterns" |
    | "a mug" | "minimalist ceramic coffee mug, matte glaze, slight imperfections" |
    | "skin" | "weathered, sun-damaged skin, deep expression lines" |
    | "metal" | "brushed steel with matte finish, fingerprint smudges" |
    
    **Surface properties:**
    - Matte / glossy / satin
    - Rough / smooth / textured
    - Translucent / opaque / refractive
    - Worn / pristine / aged / patinated
    
    ## Creative Prompt Examples
    
    **Cartoon Portrait (3D + Real):**
    ```
    Based on the uploaded reference image, create a photorealistic scene
    featuring the real human standing next to a giant 3D animation-style
    version of themselves. Both must have identical facial structures,
    clothing, and poses. Clean gray-blue studio background, cinematic lighting.
    ```
    
    **Animation → Photorealistic:**
    ```
    Convert this uploaded animated still into an ultra-realistic, cinematic
    scene. Transform the animated characters into real humans while preserving
    their original identities, facial structures, outfits, and expressions.
    ```
    
    **Historical Maps (Google Street View style):**
    ```
    Generate a hyper-realistic image of [HISTORICAL EVENT], perfectly
    replicating a Google Maps Street View capture. Include wide-angle barrel
    distortion, semi-transparent Maps UI overlay (compass, chevron arrows),
    and '© Google [YEAR]' watermark. Automatically blur faces for privacy.
    ```
    
    **Kindergarten Filter:**
    ```
    A child's crayon drawing on white lined notebook paper of [SUBJECT].
    Use chunky wax-crayon strokes, wobbly outlines, bold colors that overflow
    the lines. Visible pressure marks, waxy smudges, uneven scribble shading.
    Joyful, naive art style.
    ```
    
    **Fashion Editorial:**
    ```
    [Subject description]. [Action/pose]. [Setting].
    Medium-full shot, center-framed.
    Fashion magazine style editorial, shot on medium-format analog film,
    pronounced grain, high saturation, cinematic lighting effect.
    ```
    
    ---
    
    *Author: Serge Shima ([t.me/aimastersme](https://t.me/aimastersme) · [sergeshima.com](https://sergeshima.com) · [aimasters.me](https://aimasters.me)) · License: CC BY 4.0 — attribution required · Source: [smixs/visual-skills](https://github.com/smixs/visual-skills)*
    
  • image/references/de-slop.md7.1 KB0fbd9e29fc6aupstream ↗
    # De-Slop — killing the "AI look"
    
    Load this whenever the brief is **a real photograph**: portrait, reportage, product-in-hand, UGC, casting shot, anything that must pass as an unretouched capture. Also load it when the user complains that a result "looks AI", "too glossy", "not like the reference".
    
    Do NOT load it for illustration, 3D, poster graphics, stylized art — there the polish is the point.
    
    *De-slop* is the name of this whole discipline: casting, capture, located imperfection, grading. The booster-word hygiene inside it is the same rule GPT Image 2.5 already carries as §Anti-Slop in [gpt-image.md](gpt-image.md) — that file owns the model syntax, this one owns the discipline.
    
    The premise: both model families carry a strong prior toward an idealized image. Left alone they beautify — skin loses pores, faces gain symmetry, clothes lose creases, backgrounds gain saturation. Anything in the reference that differs from that ideal is quietly discarded unless you escalate it.
    
    ## 1. The default prior — where beautification starts
    
    Check every axis before writing. If the brief or reference differs from the middle column, the deviation must be escalated (§2).
    
    | Axis | What the model reaches for | What a real reference usually has |
    |---|---|---|
    | Eyes | large, round, wide-open, symmetric | narrow, hooded, monolid, one eye smaller |
    | Face | V-line jaw, doll proportions, mirror symmetry | long or flat midface, visible asymmetry |
    | Skin | poreless, glowing, even tone | pores, peach fuzz, redness zones, uneven tone |
    | Body | idealized silhouette | flat, square, heavy, or simply ordinary |
    | Clothing | tighter, newer, more skin | the exact coverage and wear of the reference |
    | Hair | volumized, glossy, every strand placed | flatter, clumped, flyaways, dull ends |
    | Stance | model pose, chin up | the awkward real stance from the reference |
    | Casting | idol / influencer look | an ordinary person |
    | Background | saturated, tidy, gradient | muted, cluttered, one named hex |
    | Extras | adds earrings, props, signage, text | explicitly absent |
    
    Counter-steer toward the reference, not toward plainness. If the subject genuinely is glamorous, describe that faithfully — only the capture rules in §4-§6 still apply.
    
    ## 2. Escalate every deviation in two places
    
    A deviation stated once, in passing, loses to the prior. State it twice:
    
    1. **In the descriptive slot**, geometrically and measurably — "narrow eyes with a low single lid crease, iris about 60% covered", not "unusual eyes".
    2. **In `Constraints`** (GPT Image 2.5 fifth slot, see [gpt-image.md](gpt-image.md)) — one imperative line per axis, labeled: `EYES:`, `SKIN:`, `BODY:`, `HAIR:`, `MEDIUM:`.
    
    This is the one place where naming the unwanted version is legitimate — the `Constraints` slot exists for "what must NOT change or appear". Everywhere else keep the positive framing from [golden-rules.md](golden-rules.md) §2.
    
    Nano Banana has no constraints slot: fold the same lines into one short closing sentence of the prose prompt.
    
    ## 3. Boosters banned from positive slots
    
    These words push the prior harder, they do not add fidelity: `4K`, `8K`, `ultra-detailed`, `hyper-detailed`, `ultra-realistic`, `photorealistic render`, `masterpiece`, `best quality`, `sharp focus`, `crisp`, `intricate detail`, `flawless`, `stunning`, `perfect`.
    
    On GPT Image 2.5 they actively degrade the result ([gpt-image.md](gpt-image.md) §Anti-Slop). On Nano Banana they are mostly inert, but they crowd out the facts that would have worked.
    
    ## 4. A capture pipeline instead of adjectives
    
    Do not ask for "realistic". Name the device and the path the file took — that is what carries the texture.
    
    | Medium | What to state | Residue to ask for |
    |---|---|---|
    | Video frame grab | grabbed from handheld footage, not a photo | soft detail, mild compression blocking, slight motion smear |
    | Film scan | 35mm colour negative, lab scan | visible grain, halation around highlights, soft blacks |
    | Phone photo | phone camera, auto HDR off | smeared shadow noise, faint edge halo, slightly flat tone |
    | Mirrorless / DSLR still | natural light, no flash, no retouch | fine sensor noise, natural falloff, unpolished skin |
    
    State the medium as a fact of the image, never as a render style: "a real phone photo", not "photorealistic render in phone style".
    
    ## 5. Imperfections, and where they sit
    
    An imperfection without a location gets ignored or sprayed everywhere. Name the place:
    
    - skin — "pores across the nose and cheeks, redness at the nostril wings, one mole below the left jaw"
    - hair — "two flyaways above the crown, ends clumped on the right shoulder"
    - fabric — "creases at the elbow bend, collar sitting slightly askew on the left"
    - environment — "scuffed baseboard behind the subject, uneven paint on the wall"
    - optics — "one blown highlight on the forehead, slight motion blur on the near hand"
    
    Calibrate to the reference, but never set the count to zero. See the micro-detail checklist in [prompt-framework.md](prompt-framework.md) §Cinematic Verbose for the same discipline applied to hero shots.
    
    ## 6. Sharpness and grading
    
    - Ask for no added sharpening and low micro-contrast. Never write "sharp focus" or "crisp" (see §3).
    - Grading: flat or natural, muted-to-moderate saturation, low contrast — unless the reference clearly shows otherwise.
    - Ban the glow family in `Constraints`: HDR, bloom, high clarity, luminous or radiant skin, glossy hair highlights, perfectly even background.
    - On GPT Image 2.5 stay at `quality: medium` for this kind of work. The table in [gpt-image.md](gpt-image.md) §Quality Settings sends portraits to `high` — that entry optimizes identity fidelity, which here fights the brief: `high` sharpens micro-texture until it reads as retouching. Step up to `high` only when matching a supplied face outranks the unretouched look, then re-run this section's checks on the result.
    
    ## 7. Self-check before sending
    
    - Every axis in §1 that deviates appears both in a descriptive slot and in `Constraints`.
    - No booster word from §3 survives in any positive slot.
    - A capture pipeline from §4 is named, and the medium is stated as fact, not as style.
    - At least three located imperfections from §5, each with a place.
    - Sharpness and grading clauses present (§6).
    - Anything the model likes to add but the brief does not contain — glasses, earrings, a second person, signage, captions — is listed as absent.
    - Nothing in `Constraints` contradicts a positive slot (asking for grain while banning grain).
    
    ## 8. When the result still misses the reference
    
    Treat every mismatch as a variable that was under-specified, not as a bad roll. Split the offending description into finer sub-facts, add or sharpen its `Constraints` line, and re-run the same prompt — one change per iteration, per [golden-rules.md](golden-rules.md) §3. Re-rolling an unchanged prompt returns the prior.
    
    ---
    
    *Author: Serge Shima ([t.me/aimastersme](https://t.me/aimastersme) · [sergeshima.com](https://sergeshima.com) · [aimasters.me](https://aimasters.me)) · License: CC BY 4.0 — attribution required · Source: [smixs/visual-skills](https://github.com/smixs/visual-skills)*
    
  • image/references/dimensional.md2.2 KB03c107bcff44upstream ↗
    # Dimensional Translation
    
    2D↔3D conversions. Floor plans to renders, memes to 3D.
    
    ## 2D Floor Plan → 3D
    
    ```
    Based on this 2D floor plan, generate interior visualization.
    
    Layout: Design board format
    Main image: Wide-angle [ROOM] perspective (top 60%)
    Below: 3 smaller renders - [ROOM A], [ROOM B], [ROOM C]
    
    Style: [Modern minimalist / Scandinavian / Industrial]
    Materials: [flooring], [walls], [textiles]
    Lighting: Soft natural light, large windows
    Quality: Photorealistic rendering
    ```
    
    ## Sketch → 3D Concept
    
    ```
    Convert this sketch into [STYLE] visualization.
    Perspective: [CAMERA ANGLE]
    Lighting: [SETUP]
    Quality: [photorealistic / stylized / rendered]
    ```
    
    ## 3D → 2D Technical
    
    ```
    Flatten this 3D scene into [STYLE].
    Views: [plan / elevation / section]
    Labels: [ANNOTATIONS]
    Line work: technical, precise
    ```
    
    ## Meme Conversion
    
    ```
    Turn "[MEME NAME]" meme into [NEW STYLE].
    Keep composition identical.
    Character: [plush toy / realistic / cartoon]
    Environment: [realistic flames / stylized / etc]
    ```
    
    Example:
    - "Turn 'This is Fine' dog into photorealistic 3D render"
    - "Make the dog look like plush toy, fire realistic"
    
    ## Architectural Visualization
    
    **Exterior:**
    ```
    Generate exterior visualization from this elevation.
    Time: [dawn / midday / dusk / night]
    Weather: [clear / overcast / dramatic clouds]
    Context: [landscaping / street / aerial view]
    ```
    
    **Interior:**
    ```
    Generate interior from this plan.
    View: [living area / kitchen / bedroom]
    Style: [AESTHETIC]
    Lighting: [natural / artificial / mixed]
    Props: [furniture / decor / people]
    ```
    
    ## Product 2D → 3D
    
    ```
    Convert this flat design into 3D product mockup.
    Material: [plastic / metal / fabric / glass]
    Lighting: [studio / environmental]
    Angle: [front / 3-quarter / top-down]
    ```
    
    ## Quality Levels
    
    - "Sketch quality" - rough, conceptual
    - "Presentation quality" - clean, professional
    - "Photorealistic" - indistinguishable from photo
    - "Stylized render" - artistic interpretation
    
    ---
    
    *Author: Serge Shima ([t.me/aimastersme](https://t.me/aimastersme) · [sergeshima.com](https://sergeshima.com) · [aimasters.me](https://aimasters.me)) · License: CC BY 4.0 — attribution required · Source: [smixs/visual-skills](https://github.com/smixs/visual-skills)*
    
  • image/references/editing.md2.9 KB2a36fbd90d29upstream ↗
    # Image Editing
    
    Универсальные паттерны. Логика edit'а различается по моделям:
    
    - **Nano Banana:** conversational, без масок. «Keep X same, change Y». Хорошо понимает физику и материалы.
    - **GPT Image 2.5:** двухколоночная логика **Change / Preserve / Constraints**. Preserve list повторять каждую итерацию, иначе drift. Опционально `mask_image_url` для точечных edit'ов. Лучшая identity preservation. См. [gpt-image.md](gpt-image.md#editing--двухколоночная-логика).
    
    Общее правило: **один edit за итерацию**, не пытайся менять всё разом.
    
    ## Object Removal
    ```
    Remove [OBJECT] from this image.
    Fill with [LOGICAL REPLACEMENT] matching surroundings.
    Keep [PRESERVED ELEMENTS] exactly the same.
    ```
    
    Examples:
    - "Remove tourists, fill with cobblestones"
    - "Remove car, extend street naturally"
    
    ## Object Addition
    ```
    Add [OBJECT] to this image.
    Position: [LOCATION]
    Style: matching existing lighting
    Scale: [RELATIVE SIZE]
    ```
    
    ## Lighting Control
    ```
    Change lighting to [NEW LIGHTING].
    Keep subject and composition same.
    ```
    
    Vocabulary:
    - Golden hour / sunset / warm backlight
    - Overcast / soft / diffused
    - Night / dramatic / single source
    - Rim lighting / silhouette
    
    ## Seasonal/Weather
    ```
    Turn scene into [SEASON/WEATHER].
    Keep architecture exactly same.
    Adjust: [snow/leaves/reflections/sky]
    ```
    
    ## Colorization
    
    **Photo:**
    ```
    Colorize this B&W photograph.
    Era-appropriate colors for [DECADE].
    Skin tones: natural, realistic
    ```
    
    **Manga:**
    ```
    Colorize this manga panel.
    Style: [vibrant anime / muted realistic]
    Effects: [glowing/neon] for energy elements
    Maintain: line art integrity
    ```
    
    ## Restoration
    ```
    Restore this damaged photograph.
    Fix: [tears/scratches/fading/stains]
    Enhance: sharpness, contrast
    Preserve: original character and grain
    ```
    
    ## Localization
    ```
    Translate all [SOURCE] text to [TARGET].
    Keep everything else same.
    Maintain: font style, sizing, position
    ```
    
    Cultural adaptation:
    ```
    Localize this [ORIGINAL] ad to [TARGET MARKET].
    Background: [NEW LOCATION]
    Translate: text to [LANGUAGE]
    Keep: brand elements, core composition
    ```
    
    ## Physics-Aware
    
    NBP understands materials:
    ```
    Fill this glass with [LIQUID].
    Add: refraction, meniscus, condensation
    Match: existing lighting
    ```
    
    ```
    Add [MATERIAL] texture to [SURFACE].
    Properties: [matte/glossy], [rough/smooth]
    ```
    
    ## Conversational Refinement
    
    After initial edit:
    - "Make it warmer"
    - "Increase contrast"
    - "Soften the edges"
    - "Add more detail to shadows"
    
    ---
    
    *Author: Serge Shima ([t.me/aimastersme](https://t.me/aimastersme) · [sergeshima.com](https://sergeshima.com) · [aimasters.me](https://aimasters.me)) · License: CC BY 4.0 — attribution required · Source: [smixs/visual-skills](https://github.com/smixs/visual-skills)*
    
  • image/references/golden-rules.md8.9 KBf61042645b80upstream ↗
    # Golden Rules
    
    Универсальные принципы. Работают для **обеих** семей моделей (Nano Banana и GPT Image 2.5).
    Для модельной специфики см. [nano-banana.md](nano-banana.md), [gpt-image.md](gpt-image.md).
    
    ## 1. Start with a Verb
    
    Tell the model the primary operation: "Create", "Generate", "Design", "Transform", "Convert", "Edit". This sets the intent before details.
    
    ## 2. Positive Framing
    
    Describe what you WANT, not what you don't want. The model understands presence better than absence.
    
    - ✅ "empty street" → ❌ "street with no cars"
    - ✅ "clean background" → ❌ "no clutter"  
    - ✅ "solo portrait" → ❌ "no other people"
    
    ## 3. Edit, Don't Re-roll
    
    Image 80% correct? Request specific change:
    - "Change lighting to sunset"
    - "Make text neon blue"
    - "Move chart to right third"
    
    ## 4. Natural Language
    
    ❌ Bad: "Cool car, neon, city, night, 8k"
    ✅ Good: "A cinematic wide shot of a futuristic sports car speeding through a rainy Tokyo street at night. Neon signs reflect off wet pavement and metallic chassis."
    
    ## 5. Be Specific
    
    | Element | Vague | Specific |
    |---------|-------|----------|
    | Subject | "a woman" | "sophisticated elderly woman in vintage Chanel-style suit" |
    | Material | "shiny" | "brushed steel with matte finish" |
    | Color | "dark green" | "#0d3d2d deep emerald" |
    | Position | "on the right" | "right third, bleeding off edge" |
    
    ## 6. Provide Context
    
    Context helps model make logical decisions:
    - "for Brazilian gourmet cookbook" → infers professional plating, shallow DOF
    - "for executive strategy presentation" → infers corporate aesthetic
    - "for children's educational app" → infers friendly, colorful style
    
    ## 7. Quote Text Exactly
    
    Any text for rendering goes in quotes:
    - "[HEADLINE TEXT]"
    - Labels: "[Revenue Growth]", "[Net Income]"
    - Specify weight: bold, thin, extra bold
    - Specify position: "upper third", "centered"
    
    ## Prompt Template
    
    ```
    Create a [TYPE] for [CONTEXT].
    
    Background: [Description with hex colors]. [Atmospheric effects].
    
    [HERO ELEMENT]:
    [Detailed description - position, lighting, angle]
    
    Typography:
    Line 1: "[TEXT]" in [weight], [color], [size], [position]
    Line 2: "[TEXT]" in [weight], [color], [size], [position]
    
    [ADDITIONAL ELEMENTS]
    
    Mood: [Emotional descriptor]
    Format: [ASPECT RATIO]
    ```
    
    > **Thinking Mode** (NB only), **`quality: low/medium/high/xhigh/max`** (GPT Image 2.5 only) — см. соответствующие references.
    
    ## Cost Optimization (Batch Work)
    
    - **Nano Banana:** прогон вариантов на `0.5K` Flash → отбор → переген победителя на `2K`/`4K`.
    - **GPT Image 2.5:** прогон на `quality: low` → отбор → переген на `medium` или `high`.
    
    В обоих случаях: дешёвая разведка → дорогой финал.
    
    ## Conversational Refinement
    
    After generation:
    - "Change the headline color to #3b82f6"
    - "Add subtle drop shadow to text"
    - "Increase contrast, make it more dramatic"
    - "Soften the background, add blur"
    
    ## Reference Images
    
    Multi-image вход: **NB до 14**, **GPT Image 2.5 до 16**. Индексируй с ролью каждой картинки.
    
    Используй для:
    
    **Вписать в существующий дизайн:**
    ```
    [Attach design/layout image]
    Create content following this exact layout and style.
    Replace [ELEMENT] with [NEW CONTENT].
    Keep colors, typography, composition.
    ```
    
    **Лицо/персонаж как референс:**
    ```
    [Attach portrait]
    Use this person's face. Keep features exactly the same.
    Change: [expression/pose/setting]
    ```
    
    **Продукт/объект как референс:**
    ```
    [Attach product photo]
    Place this product in [NEW CONTEXT].
    Match lighting and perspective.
    ```
    
    **Стиль как референс:**
    ```
    [Attach style reference]
    Create [NEW CONTENT] in this exact visual style.
    Match colors, textures, mood.
    ```
    
    **Несколько референсов сразу:**
    ```
    [Attach Image 1 - face]
    [Attach Image 2 - outfit]
    [Attach Image 3 - background]
    Combine: face from Image 1, outfit style from Image 2, setting from Image 3.
    ```
    
    ## World Knowledge Anchors
    
    GPT Image 2.5 обладает глубокими знаниями о культуре, эпохах и визуальных стилях. Вместо описания каждой детали — дай модели культурный/временной/жанровый якорь, и она сама заполнит аутентичные детали.
    
    ### Три типа якорей
    
    **Era anchors** — временной и географический маркер, который вызывает целый визуальный мир:
    - "Bethel, NY, August 1969" → Woodstock aesthetic без необходимости описывать тай-дай, грязь, сцены
    - "Berlin, November 1989" → падение стены, толпы, граффити, эйфория
    - "Tokyo, 1982" → неоновый Shinjuku, аналоговая электроника, ранний cyberpunk
    
    **Cultural anchors** — перенос визуального языка одного культурного объекта на другой контекст:
    - "{game_title} in {real_city}" → автоматически применяет визуальный стиль игры к реальной локации (GTA style, Persona style, и т.д.)
    - "Soviet constructivism poster about {modern_topic}" → стиль Родченко/Эль Лисицкого на современную тему
    - "Ukiyo-e print of {modern_scene}" → японская гравюра с современным содержанием
    
    **Genre anchors** — режиссёр/фотограф/движение как линза:
    - "Peter Lindbergh influence" → сильный Ч/Б, минимальная ретушь, raw editorial
    - "Wes Anderson palette" → симметричный кадр, пастельная палитра, центрированная композиция
    - "Studio Ghibli mood" → мягкое акварельное небо, зелёная листва, тёплый ностальгический свет
    - "Roger Deakins lighting" → натуральный свет, глубокие тени, кинематографичный объём
    
    ### Правила использования
    
    1. **Используй как HIGH-LEVEL steering** — якорь задаёт настроение и эстетику, а не заменяет весь промпт
    2. **Комбинируй с конкретными визуальными деталями** — якорь устанавливает мир, детали устанавливают специфику
    3. **Не стакай несколько genre anchors** — выбери один. "Peter Lindbergh + Wes Anderson" = каша
    4. **Era/cultural anchors работают лучше с GPT Image 2.5** (world knowledge). С Nano Banana результат менее предсказуем — NB больше опирается на явные описания
    
    ### Примеры
    
    **Era anchor + конкретные детали:**
    ```
    Create an editorial portrait set in Havana, 1957.
    
    Subject: jazz musician leaning against pastel-colored colonial building,
    holding trumpet loosely at his side. Linen suit, open collar.
    Lighting: harsh Caribbean afternoon sun, deep shadows under awning.
    Format: 3:4
    ```
    > "Havana, 1957" вызывает: старые американские машины на заднем плане, облупившаяся штукатурка, кованые балконы, тропическая атмосфера — без необходимости это описывать.
    
    **Cultural anchor + новый контекст:**
    ```
    Create a scene of a quiet Kyoto temple garden, rendered in the visual style
    of Studio Ghibli. Morning mist over moss-covered stones, a single monk
    sweeping fallen maple leaves. Soft watercolor textures, warm nostalgic palette.
    Format: 16:9
    ```
    > "Studio Ghibli" задаёт акварельность, теплоту, ностальгию. Детали (мох, клён, монах) задают конкретную сцену.
    
    **Genre anchor + специфика:**
    ```
    Create a fashion editorial portrait with Peter Lindbergh influence.
    
    Subject: model in oversized men's blazer, no makeup, wind-tousled hair.
    Setting: empty winter beach, overcast sky.
    Mood: raw, unpolished beauty
    Format: 2:3
    ```
    > "Peter Lindbergh influence" даёт: мощный Ч/Б (или desaturated), отсутствие ретуши, raw emotional quality. Детали (пляж, блейзер, ветер) конкретизируют кадр.
    
    ---
    
    *Author: Serge Shima ([t.me/aimastersme](https://t.me/aimastersme) · [sergeshima.com](https://sergeshima.com) · [aimasters.me](https://aimasters.me)) · License: CC BY 4.0 — attribution required · Source: [smixs/visual-skills](https://github.com/smixs/visual-skills)*
    
  • image/references/gpt-image.md14 KB96d600dfa337upstream ↗
    # GPT Image 2.5 — Specific Rules
    
    Актуальное поколение OpenAI (сентябрь 2026). Две модели, единый API и единые правила промптинга:
    
    | Модель | Model ID | Роль |
    |--------|----------|------|
    | **GPT Image 2.5 Flare** | `gpt-image-2.5-flare` | Default. Малая модель, оптимизирована на скорость; качество выше GPT Image 2 при ~50% меньшей латентности |
    | **GPT Image 2.5 Sunburst** | `gpt-image-2.5-sunburst` | Базовая модель, оптимизирована на качество: precision-edits, identity preservation, кампейн-креатив, продуктовая съёмка |
    
    Выбор: рутинная генерация и высокие объёмы → Flare; сложные многошаговые edits, критичное сохранение лиц/продуктов/лейблов → Sunburst. Рабочий паттерн: черновики на Flare, финальные деливераблы прогонять на Sunburst только там, где Flare не дотянул. GPT Image 2 и старше — migration-only, в новых промптах не использовать.
    
    ## Структура промпта — 5 slots
    
    GPT Image 2.5 сильнее всего реагирует на разделение по секциям. Для сложных запросов пиши лейблами, не сплошным текстом (офф. дока: «organize the prompt as scene, subject, details, and constraints, using labeled sections»). Формат при этом свободный — короткий промпт, параграф, JSON-подобная структура или теги работают одинаково; выбирай тот, который проще читать и обновлять.
    
    ```
    Scene: location, time of day, background, environment
    Subject: primary focus — who or what is central
    Important Details: materials, textures, lighting, camera angle, mood, composition
    Use Case: editorial, product mockup, UI, poster, infographic
    Constraints: what must NOT change/appear (no watermarks, preserve face, no extra text)
    ```
    
    > «The fifth slot is where most mediocre prompts fail silently.» Без явных constraints модель дрейфует.
    
    ### Минимальный пример
    
    ```
    Scene: small Lisbon florist storefront at blue hour, wet cobblestones
    Subject: woman in navy apron locking the front door, half-turned to camera
    Important Details: warm interior glow spilling onto pavement, 50mm feel,
      soft contact shadows, brushed brass door handle, "Florista" hand-painted sign
    Use Case: editorial photography
    Constraints: no extra signage, no people in background, no text other than the sign
    ```
    
    ## Anti-Slop Rules
    
    GPT Image 2.5 особенно чувствителен к качеству формулировок. Vague praise = деградация результата.
    
    | ❌ Не пиши | ✅ Пиши |
    |-----------|---------|
    | stunning, incredible, epic, gorgeous, masterpiece | overcast daylight, brushed aluminum, chipped paint, 50mm feel |
    | «minimalist brutalist luxury photoreal» (стиль-теги) | «cream background, heavy black sans-serif, asymmetrical type block, one hero object, generous negative space» |
    | «как в Apple-рекламе» | конкретные визуальные факты |
    | мудовый язык, в котором тонут функциональные требования | прямое заявление: «image must contain a transit kiosk» |
    
    ## Quality Settings — пятиступенчатая шкала
    
    `quality`: `low` · `medium` · `high` · `xhigh` · `max` (+ `auto`). Это **новая шкала**, не переименование трёх старых уровней GPT Image 2 — не переноси настройки один-в-один.
    
    | Setting | Когда |
    |---------|-------|
    | `low` | High-volume, превью, exploratory, latency-sensitive, draft |
    | `medium` | **Default starting point** |
    | `high` | Маленький/плотный текст, infographics, портреты, identity-sensitive edits, brand assets |
    | `xhigh` | Тонкие текстуры, насыщенные детальные сцены, print-ready ассеты |
    | `max` | Максимальная fidelity модели — только если `xhigh` не закрыл конкретное требование |
    
    Методика офф. доки: стартуй с `medium`; не хватает — шаг вверх; хватает — проверь шаг вниз ради latency. `xhigh`/`max` только когда они закрывают конкретное невыполненное требование в рамках latency-бюджета. Меняй один параметр за раз: сначала quality, потом переписывание промпта.
    
    ## Размеры
    
    - Max edge: ≤3840px
    - Обе стороны: кратны 16
    - Aspect ratio: max 3:1 (long:short)
    - Total pixels: 655 360 – 8 294 400
    - `size: auto` или кастомный `WIDTHxHEIGHT`
    
    **Ходовые (из офф. доки):**
    - Portrait 1024×1536 · Landscape 1536×1024 · Square 1024×1024
    - 2K: 2048×2048, 2048×1152
    - 4K: 3840×2160 (landscape), 2160×3840 (portrait)
    
    **`background`:** `auto` | `opaque` | `transparent` (для transparent — PNG или WebP, проверяй реальный alpha-канал, не «нарисованный» фон).
    
    > Экстрим вроде 1:8 / 8:1 GPT Image 2.5 НЕ умеет (лимит 3:1) — иди в Nano Banana.
    
    ## Text in Image
    
    - Литеральный текст в `"..."` или ALL CAPS.
    - Укажи, **сколько раз** текст должен появиться («render the tagline exactly once»).
    - Шрифт, размер, цвет, позиция — явно.
    - Сложные слова и бренды: спеллинг по буквам.
    - Защита от мусора: «**no extra text, no duplicate text, no watermarks**».
    - Для нечитаемого мелкого текста: «**100 percent readable and physically believable**».
    - Маленький/плотный/multi-font → `quality: high` минимум; print-ready плотная типографика → `xhigh`.
    
    ## Editing — двухколоночная логика
    
    Endpoint: `images.edit` (OpenAI API) или `openai/gpt-image-2.5-flare/edit` · `openai/gpt-image-2.5-sunburst/edit` (fal/wavespeed-провайдеры). Для edit-heavy пайплайнов приоритет Sunburst.
    
    ```
    Change: [single concrete change]
    Preserve: face, identity, pose, lighting, framing, background, geometry, text, layout
    Constraints: no extra objects, no redesign, no drift
    ```
    
    **Правила edit:**
    - **Один edit за итерацию.** Не пытайся менять всё разом.
    - **Preserve list повторять каждую итерацию.** «Same style as before» несёт контекст, но при drift — restate critical constraints.
    - **Surgical edits:** явно перечисли что НЕ трогать (saturation, contrast, layout, arrows, labels, camera angle).
    - Input fidelity всегда high — отдельного параметра нет.
    - Опционально: маска для точечных edits (см. editing with a mask в офф. доке).
    
    ### Edit-паттерны
    
    **Virtual try-on:** «Change garments only. Preserve exact face, body shape, pose, hair, expression, background, camera angle. Match lighting/shadows so outfit looks naturally worn.»
    
    **Object removal:** «Remove [X]. Do not change anything else.»
    
    **Lighting/weather swap:** «Change ONLY environmental conditions: lighting direction/quality, shadows, atmosphere, precipitation. Preserve identity, geometry, camera angle, object placement.»
    
    **Interior swap:** «Swap [furniture]. Preserve camera angle, lighting, shadows, surrounding context. Photorealistic contact shadows.»
    
    **Translate in place:** «Translate the text in the infographic to [lang]. Do not change any other aspect of the image.» — потом проверь перевод и слова, оставшиеся на исходном языке.
    
    **Sketch-to-render:** «Turn this drawing into a photorealistic image. Preserve the exact layout, proportions, and perspective. Choose realistic materials and lighting consistent with the sketch intent. Do not add new elements or text.»
    
    ## Multi-Image — до 16 рефов
    
    Индексируй с **ролью**, не только номером (subject / style / clothing / background):
    ```
    Image 1: base scene
    Image 2: jacket reference (apply only the jacket fabric/cut to subject in Image 1)
    Image 3: lighting reference (apply golden-hour quality from Image 3)
    ```
    
    ## Style Transfer
    
    Не пиши абстрактно («minimalist», «editorial»). Назови конкретные визуальные свойства референса: палитра, edge treatment, силуэт, обработка теней, plane логика. Референсу — явную роль: «use the palette and texture from the input image».
    
    ## World Knowledge
    
    GPT Image 2.5 умеет домысливать контекст: «Bethel, NY, August 1969» → выведет Woodstock-эстетику. Используй: дай исторический/культурный анкер, не расписывай каждую деталь — но проверяй одежду, стейджинг и окружение на историческую точность.
    
    ## Iteration Strategy
    
    - Стартуй с **чистого** базового промпта.
    - Один change за раунд. «Make lighting warmer», «remove extra tree», «restore original background».
    - При drift — перечисли invariants заново.
    - Для длинных промптов — labeled sections, не одна простыня.
    - Character consistency в серии: создай character reference одним прогоном, дальше переиспользуй картинку как input и **повторяй defining details персонажа в каждом промпте**.
    
    ## Check the Result (чек-лист офф. доки)
    
    - Текст точен и читаем? Лейблы и связи на диаграммах корректны?
    - Identity, форма продукта, лейблы, детали референсов не уплыли?
    - Edit изменил только запрошенное?
    - Если нужна прозрачность — в файле реальный alpha-канал, а не закрашенный фон?
    
    ## Use-Case Templates
    
    ### Photoreal Editorial
    ```
    Scene: [location, time, weather]
    Subject: [who, action, framing]
    Important Details: [lens feel, light source, surface wear, imperfections, real texture]
    Use Case: editorial photograph, looks like a real photo
    Constraints: no glamorization, no heavy retouching, no studio gloss
    ```
    
    ### Product Mockup (Clean Background)
    ```
    Scene: plain white opaque background
    Subject: [product] centered
    Important Details: crisp silhouette, no halos/fringing, light contact shadow,
      preserve label legibility exactly, preserve geometry
    Use Case: product mockup
    Constraints: no restyling, only background removal + light polish
    ```
    
    ### UI Mockup
    ```
    Scene: [device frame, e.g. iPhone 15 Pro]
    Subject: [screen/app name] — describe AS IF IT EXISTS, not concept art
    Important Details: layout, hierarchy, real interface elements, exact copy in quotes,
      typography behavior, spacing, state
    Use Case: shipped product screenshot
    Constraints: no sketch language, no placeholder text, no Lorem Ipsum
    Quality: high (for small UI text)
    ```
    
    ### Marketing Creative with Text
    ```
    Scene: [environment]
    Subject: [hero element]
    Important Details: [composition, palette, mood]
    Use Case: ad creative for [audience]
    Text: "EXACT HEADLINE" in [font style], [color], [position], exactly once
          "exact subhead" in [font style], [color], [position]
    Constraints: no extra text, no duplicate text, no watermarks, no unrelated logos
    Quality: high
    ```
    
    ### Infographic / Diagram
    ```
    Title: "[TITLE]"
    Content flow: [step 1] → [step 2] → [step 3]
    Visual format: [layout type — flowchart, pyramid, isometric, etc.]
    Use Case: educational infographic for [audience]
    Constraints: readable labels at all sizes, clear hierarchy, no clutter,
      no decorative noise, ample whitespace
    Quality: high
    Size: 1536×1024
    ```
    
    ### Logo (transparent)
    ```
    Subject: [brand] logo — [defining shapes], flat design, minimal strokes,
      no gradients unless essential, legible at small and large sizes
    Use Case: reusable brand mark
    Constraints: single centered logo, generous padding, clean alpha edges,
      no solid backdrop, no scenery, no checkerboard, no watermark
    Background: transparent · Format: png
    ```
    
    ## Migration from GPT Image 2 / 1.5 / 1
    
    - Промпты, референсы, размеры переносятся как есть — первый прогон делай без изменений и сравни.
    - `quality` не мапится 1:1: шкала новая. Реши, что важнее — та же fidelity (старый `medium` ≈ новый `high`) или то же имя тира дешевле.
    - `input_fidelity` (был в 1.5/1) убрать — high всегда.
    - Стартовая точка: Sunburst для качества → если проходит, проверь Flare ради latency; либо сразу Flare, если GPT Image 2 уже устраивал.
    - Смотри instruction following, identity/product preservation, точность текста, unwanted changes, transparency; повторяй запросы для оценки консистентности. Для editing-пайплайнов тестируй всю цепочку edits, не только шаги по отдельности.
    
    ---
    
    *Author: Serge Shima ([t.me/aimastersme](https://t.me/aimastersme) · [sergeshima.com](https://sergeshima.com) · [aimasters.me](https://aimasters.me)) · License: CC BY 4.0 — attribution required · Source: [smixs/visual-skills](https://github.com/smixs/visual-skills)*
    
  • image/references/models.md6.9 KB6702f01ec664upstream ↗
    # Model Selection — Nano Banana vs GPT Image 2.5
    
    Skill пишет промпты под две семьи моделей. Они мыслят по-разному — выбор модели меняет структуру промпта.
    
    ## TL;DR
    
    | Задача | Модель |
    |--------|--------|
    | Реальное место/объект (с грунтингом) | **Nano Banana** (NB2/NBP) |
    | Сложная сцена с физикой/композицией | **Nano Banana Pro** |
    | Длинные горизонтальные/вертикальные форматы (1:8, 8:1, 4:1) | **Nano Banana** (только NB поддерживает экстрим) |
    | Дешёвая массовая генерация | **Nano Banana 2 Lite** или **GPT Image 2.5 Flare** (`quality: low`) |
    | Фотореализм с тонкой типографикой/UI | **GPT Image 2.5** |
    | Точное editing с preservation (try-on, swap, weather) | **GPT Image 2.5** (в editing у него лучшая identity-preservation) |
    | Маленький плотный текст в кадре | **GPT Image 2.5** (`quality: high`) |
    | Брендовая полиграфия / постеры с EXACT TEXT | **GPT Image 2.5** |
    | Сториборды, комиксы (последовательность) | **Nano Banana** (extreme ratios + thinking) |
    | Storyboard с фокусом на типографике | **GPT Image 2.5** |
    | Style transfer без упоминаемых референс-картинок | **GPT Image 2.5** (concrete visual targets) |
    | Рендер из 14+ референсов | **Nano Banana Pro** (до 14) или **GPT Image 2.5** (до 16) |
    
    > Внутри GPT Image 2.5 две модели: **Flare** (`gpt-image-2.5-flare`) — быстрый default, качество выше GPT Image 2 при ~50% меньшей латентности; **Sunburst** (`gpt-image-2.5-sunburst`) — базовая, для precision-edits и кампейн-качества. Промпт-правила одинаковые. GPT Image 2 — migration-only.
    
    ## Когда что выигрывает
    
    ### Nano Banana выигрывает в
    - **Image grounding.** NB2 ищет реальные изображения в интернете перед генерацией — точная архитектура конкретного храма, моста, площади; конкретные виды животных, растений. GPT Image 2.5 этого не делает.
    - **Экстремальные пропорции.** 1:8, 8:1, 1:4, 4:1 — баннеры, скроллы, комикс-стрипы. У GPT Image 2.5 max 3:1.
    - **«Thinking» режим.** Сложные инфографики со spatial logic.
    - **Цена/скорость.** NB2 = $0.04/img.
    
    ### GPT Image 2.5 выигрывает в
    - **Identity preservation в edit.** Меняешь одежду / погоду / фон — лицо, поза, геометрия не плывут. Двухколоночная логика (change / preserve) работает как контракт.
    - **Тонкий текст в кадре.** Маленькие подписи, легенды, footnotes, multi-font layouts. На `quality: high` рендерит чётче.
    - **UI-моки и продуктовые скриншоты.** Иерархия, реальные интерфейс-элементы, читаемые лейблы.
    - **Структурированный 5-slot промпт.** Чёткое разделение Scene/Subject/Details/Use case/Constraints даёт предсказуемость.
    - **`quality` рычаг.** low/medium/high/xhigh/max — осознанный trade-off скорости и точности (шкала новая, пятиступенчатая; medium = default).
    - **4K-выход.** До 3840×2160 / 2160×3840 из коробки.
    
    ### Где обе модели одинаково хороши
    - Photorealistic портреты.
    - Product shots на нейтральном фоне.
    - Минималистичные постеры.
    - Editorial-фотография.
    
    ## Различия в синтаксисе промпта
    
    | Аспект | Nano Banana | GPT Image 2.5 |
    |--------|-------------|-------------|
    | Стиль промпта | Натуральный язык, 1-2 параграфа | 5-slot с лейблами секций |
    | Камера/линза | **Не указывать** числа (50mm, f/2.8) — NB игнорит | Можно «50mm feel», но как high-level look |
    | «Stunning/epic/masterpiece» | Игнорит, не вредит | **Anti-slop**: вредит, делает результат хуже |
    | Text in image | `"..."` в кавычках, font + position | `"..."` или ALL CAPS + «no extra words / no duplicate text» |
    | Negative framing | Использовать позитив | Использовать позитив + явный preserve list |
    | Сложные сцены | JSON для 5+ элементов | 5-slot template со секциями |
    | Edit | «Keep X same, change Y» | «Change: X / Preserve: Y / Constraints: Z» — повторять preserve каждую итерацию |
    | Множественные референсы | До 14, индексировать | До 16, индексировать с ролью («Image 1: base», «Image 2: jacket reference») |
    
    ## Стоимость (ориентир)
    
    | Модель | Цена | Заметки |
    |--------|------|---------|
    | Nano Banana 2 Lite | ~$0.034/img | Только 1K, ~4 сек. Черновики и массовые батчи |
    | Nano Banana 2 (Flash) | ~$0.04/img | Default для большинства задач |
    | Nano Banana Pro | ~$0.15/img | Сложные сцены, до 14 рефов |
    | GPT Image 2.5 (`low`) | ~$0.01/img (1K) | Latency-sensitive, превью, массовые батчи |
    | GPT Image 2.5 (`medium`) | ~$0.024/img (1K) | Default для GPT Image |
    | GPT Image 2.5 (`high`) | ~$0.09/img (1K) | Маленький текст, brand-sensitive, photorealism |
    | GPT Image 2.5 (`xhigh`) | ~$0.16/img (1K) | Тонкие текстуры, print-ready |
    | GPT Image 2.5 (`max`) | ~$0.36/img (1K) | Максимальная fidelity, только по конкретной нужде |
    
    Цены Flare и Sunburst одинаковые (провайдерский ориентир на 1K; 2K/4K дороже). Каждый тир дешевле одноимённого тира GPT Image 2.
    
    > Скилл сам не запускает генерацию — выдаёт промпт. Модель/quality указываем рядом с промптом как мета.
    
    ---
    
    *Author: Serge Shima ([t.me/aimastersme](https://t.me/aimastersme) · [sergeshima.com](https://sergeshima.com) · [aimasters.me](https://aimasters.me)) · License: CC BY 4.0 — attribution required · Source: [smixs/visual-skills](https://github.com/smixs/visual-skills)*
    
  • image/references/multi-panel.md18 KB8e3ac3b160b3upstream ↗
    # Multi-Panel Compositions
    
    Single-image layouts containing multiple frames, panels, or grid cells. Unlike [storyboards.md](storyboards.md) (sequential images generated one-at-a-time), these patterns produce **one image** with all panels baked in.
    
    Core principle: explicitly number and describe every panel. Models need panel-by-panel instructions — vague requests like "show multiple angles" produce inconsistent grids.
    
    ---
    
    ## 1. 9-Cell Grid Storyboard
    
    > For a product-focused narrative variant, see [patterns/ecommerce.md](patterns/ecommerce.md#9-panel-tvc-storyboard-grid).
    
    **When to use:** TVC or commercial shot breakdown — one image = 9 panels with scene titles and timestamps.
    
    ### Template (5-slot format)
    
    ```
    Scene: Dark storyboard layout — 3x3 grid of cinematic frames for a {duration}-second {product_name} commercial. Each cell has a thin dark border. Below each cell: scene number, title, and timestamp in small white sans-serif text on dark background.
    
    Subject: {product_name} commercial storyboard showing the complete narrative arc across 9 panels.
    
    Important Details:
    Panel 1 (0:00–{t1}): {scene_1_description}
    Panel 2 ({t1}–{t2}): {scene_2_description}
    Panel 3 ({t2}–{t3}): {scene_3_description}
    Panel 4 ({t3}–{t4}): {scene_4_description}
    Panel 5 ({t4}–{t5}): {scene_5_description}
    Panel 6 ({t5}–{t6}): {scene_6_description}
    Panel 7 ({t6}–{t7}): {scene_7_description}
    Panel 8 ({t7}–{t8}): {scene_8_description}
    Panel 9 ({t8}–{duration}): {scene_9_description}
    
    Each panel labeled below: "Scene {N}: {title}" and "{start}–{end}"
    
    Use Case: Pre-production storyboard for video production team
    Constraints: All 9 panels must be clearly separated, no merged cells, every panel must contain distinct content, timestamps must be legible, {language} titles
    ```
    
    **Variables:**
    - `{product_name}` — brand/product
    - `{duration}` — total spot length (e.g., "30", "60")
    - `{t1}` through `{t8}` — timestamp boundaries
    - `{scene_1_description}` through `{scene_9_description}` — shot description per panel
    - `{language}` — "Chinese" / "English" / "bilingual"
    
    **Recommended size:** 1536x1024 (landscape)
    **Model:** GPT Image 2.5 `quality: high` (text-heavy — timestamps and titles need legibility)
    **Common pitfalls:**
    - Omitting panel numbers causes the model to merge or skip panels
    - Vague scene descriptions produce near-identical panels — each must have a distinct action, angle, or subject
    - Timestamps in very small text need `quality: high` to remain readable
    
    ---
    
    ## 2. 2x2 Editorial Portrait Grid
    
    **When to use:** Same person shown from 4 angles/crops in one image for an editorial or casting look.
    
    ### Template (5-slot format)
    
    ```
    Scene: 2x2 grid of four editorial portraits, minimal gap between panels. {background_description}. {lighting_style}.
    
    Subject: {subject_description} — same person in all four panels, consistent identity and wardrobe.
    
    Important Details:
    Top-left: front-facing portrait, direct eye contact, shoulders up
    Top-right: extreme macro close-up of face — eyes, skin texture, freckles visible
    Bottom-left: lower angle looking up, chin slightly raised, confident expression
    Bottom-right: side profile, clean silhouette against background
    
    Consistent lighting across all four panels. Skin texture realistic, no airbrushing.
    
    Use Case: editorial photography, model portfolio, casting composite
    Constraints: same person in every panel, no wardrobe changes between panels, no extra people, no text overlays
    ```
    
    **Variables:**
    - `{subject_description}` — age, appearance, wardrobe (e.g., "woman in her 30s, dark curly hair, white linen shirt")
    - `{background_description}` — (e.g., "soft gray studio backdrop")
    - `{lighting_style}` — (e.g., "single key light from upper left, subtle fill from right")
    
    **Recommended size:** 1024x1024 (square) or 1024x1536 (vertical for portrait emphasis)
    **Model:** GPT Image 2.5 `quality: medium` or Nano Banana Pro (both handle photorealistic portraits well)
    **Common pitfalls:**
    - Not specifying "same person" in every panel — model may generate four different people
    - Omitting which quadrant gets which angle — model arranges arbitrarily
    - Requesting too-different crops (full body + extreme close-up) creates scale inconsistency within the grid
    
    ---
    
    ## 3. 3-Panel Campaign Collage
    
    **When to use:** Hero shot + close-up + action for a campaign visual — horizontal or vertical triptych.
    
    ### Template (5-slot format)
    
    ```
    Scene: Three-panel {orientation} collage for {brand_name} campaign. Panels separated by thin white lines.
    
    Subject: {model_description} showcasing {product_description}.
    
    Important Details:
    Panel 1 (left/top): Hero wide shot — full figure of model with product, environmental context, lifestyle setting
    Panel 2 (center/middle): Close-up product detail — {product_detail_description}, tactile textures, sharp focus
    Panel 3 (right/bottom): Action shot — model using/wearing/interacting with product, candid energy, slight motion blur on extremities
    
    Consistent warm golden-hour lighting across all panels. Same model identity throughout.
    {typography_instruction}
    
    Use Case: social media campaign visual, brand lookbook
    Constraints: same person across all panels, consistent color grading, no stock-photo stiffness, {product_name} must be visible in every panel
    ```
    
    **Variables:**
    - `{orientation}` — "horizontal" (side by side) or "vertical" (stacked)
    - `{brand_name}`, `{product_name}`, `{product_description}` — brand context
    - `{model_description}` — who appears
    - `{product_detail_description}` — what the close-up shows
    - `{typography_instruction}` — e.g., `Text: "{HEADLINE}" in bold condensed white, overlaid on Panel 1 lower third` or omit for no text
    
    **Recommended size:** 1536x1024 (horizontal triptych) or 1024x1536 (vertical triptych)
    **Model:** GPT Image 2.5 `quality: medium` — if text overlay needed, use `quality: high`
    **Common pitfalls:**
    - Not specifying panel order — "hero, close-up, action" without left/center/right assignment
    - Lighting inconsistency when mixing indoor close-up with outdoor hero shot — specify unified lighting
    - Forgetting to mention subject consistency causes three different models
    
    ---
    
    ## 4. 4x3 Borderless Grid
    
    **When to use:** 12 panels telling a story or showing moods — no gaps between panels, seamless mosaic feel.
    
    ### Template (5-slot format)
    
    ```
    Scene: 4x3 borderless grid (4 columns, 3 rows) where each of the 12 panels is an independent image but panels share no borders, gaps, or dividers — edges touch seamlessly. Overall theme: {theme}.
    
    Subject: {subject_description} — maintain strong subject consistency across all 12 panels.
    
    Important Details:
    Row 1: {panel_1}, {panel_2}, {panel_3}, {panel_4}
    Row 2: {panel_5}, {panel_6}, {panel_7}, {panel_8}
    Row 3: {panel_9}, {panel_10}, {panel_11}, {panel_12}
    
    Style: {style_description}
    Mood progression: {mood_arc}
    Each panel is an independent composition but the overall grid reads as a unified artwork.
    
    Use Case: mood board, editorial spread, social media carousel preview, album artwork
    Constraints: borderless — no white lines, no black borders, no gaps between panels. Same subject identity across all panels. No text in panels.
    ```
    
    **Variables:**
    - `{subject_description}` — who/what appears across panels
    - `{theme}` — overarching concept (e.g., "solitude in a city", "four seasons of a garden")
    - `{panel_1}` through `{panel_12}` — brief shot description per cell
    - `{style_description}` — visual style (e.g., "35mm film grain, desaturated palette")
    - `{mood_arc}` — how energy changes across the grid (e.g., "calm morning to chaotic night")
    
    **Recommended size:** 1536x1024 (landscape) — gives each cell enough resolution
    **Model:** Nano Banana Pro (handles complex multi-element compositions with thinking mode; no text needed)
    **Common pitfalls:**
    - Saying "borderless" is not enough — explicitly state "no white lines, no black borders, no gaps"
    - 12 panels in one image pushes detail limits — keep per-panel descriptions short and visually distinct
    - Without explicit subject consistency instructions, each panel may feature a different person/object
    
    ---
    
    ## 5. 6-Frame Cinematic Sequence
    
    **When to use:** Fashion editorial or film-style sequence — multiple camera angles of the same scene in one image.
    
    ### Template (5-slot format)
    
    ```
    Scene: 6-frame cinematic sequence arranged in a 3x2 grid. Dark film-strip aesthetic with thin black borders. {location_description}.
    
    Subject: {subject_description}, wearing {wardrobe_description}. Same person, same outfit, same location across all 6 frames.
    
    Important Details:
    Frame 1 (top-left): Top-down bird's eye view — subject seen from directly above, full body visible against ground/floor
    Frame 2 (top-center): Low angle looking up — subject towering over camera, dramatic perspective, sky/ceiling visible
    Frame 3 (top-right): Wide isolation shot — subject small in frame, vast environment dominates, sense of scale
    Frame 4 (bottom-left): Close-up with slight tilt — face and upper body, Dutch angle, intimate intensity
    Frame 5 (bottom-center): Motion frame — subject mid-action ({motion_action}), natural motion blur on limbs
    Frame 6 (bottom-right): Grounded final — medium shot, subject at rest, direct gaze, resolving the sequence
    
    Photographer reference: {photographer_style}
    Lighting: consistent {lighting_description} across all frames
    
    Use Case: fashion editorial, film lookbook, director's shot list visualization
    Constraints: same person and wardrobe in every frame, no costume changes, consistent color grading, no text overlays
    ```
    
    **Variables:**
    - `{subject_description}` — model details
    - `{wardrobe_description}` — clothing
    - `{location_description}` — setting
    - `{motion_action}` — what the motion frame captures (e.g., "walking forward", "turning sharply", "jumping")
    - `{photographer_style}` — (e.g., "Peter Lindbergh desaturated realism", "Helmut Newton dramatic contrast")
    - `{lighting_description}` — (e.g., "overcast natural light, soft shadows")
    
    **Recommended size:** 1536x1024 (landscape, 3x2 grid)
    **Model:** GPT Image 2.5 `quality: medium` or Nano Banana Pro
    **Common pitfalls:**
    - Not naming each frame explicitly — "various angles" is too vague, the model needs per-frame instructions
    - Top-down and low-angle in the same grid confuse the model if you don't anchor each frame to a grid position
    - Motion blur instruction must be specific ("blur on hands and feet") or the model applies blur everywhere
    
    ---
    
    ## 6. Before/After Split
    
    **When to use:** Product transformation, makeover, time comparison, renovation — two states side by side.
    
    ### Template (5-slot format)
    
    ```
    Scene: Single image split into left and right halves with a {divider_style} dividing line down the center. Before/after comparison.
    
    Subject: {subject_description}
    
    Important Details:
    Left half (BEFORE): {before_description} — {before_condition}
    Right half (AFTER): {after_description} — {after_condition}
    
    The transition at the center line should feel {transition_style}. Same camera angle, same framing, same background perspective in both halves — only the subject's state changes.
    
    Use Case: {use_case}
    Constraints: identical composition and camera angle on both sides, same lighting direction, no text unless specified, the dividing line must be clearly visible
    ```
    
    **Variables:**
    - `{subject_description}` — what is being compared
    - `{before_description}` / `{before_condition}` — left side state (e.g., "faded, cracked wall with peeling paint")
    - `{after_description}` / `{after_condition}` — right side state (e.g., "freshly painted wall, smooth finish, vibrant color")
    - `{divider_style}` — "thin white line" / "subtle gradient blend" / "sharp vertical cut"
    - `{transition_style}` — "clean and abrupt" / "natural, as if wiping away"
    - `{use_case}` — "product marketing", "renovation portfolio", "skincare results"
    
    **Recommended size:** 1536x1024 (landscape — gives each half a portrait-like proportion)
    **Model:** GPT Image 2.5 `quality: medium` — for text labels ("BEFORE" / "AFTER"), use `quality: high`
    **Common pitfalls:**
    - Not specifying "same camera angle both sides" — model may show two completely different viewpoints
    - Without a visible divider, the two halves can merge into one ambiguous scene
    - Left/right assignment matters — always state which side is before and which is after
    
    ---
    
    ## 7. 12-Panel Storyboard Poster
    
    **When to use:** Full narrative in one image — 3x4 grid (3 columns, 4 rows) for animation or video pre-production.
    
    ### Template (5-slot format)
    
    ```
    Scene: 12-panel storyboard poster, 3 columns x 4 rows. Dark background with each panel in a clean rectangular frame. "{title}" in bold white text at the top of the image. Below each panel: scene number and one-line description in small white text.
    
    Subject: {character_description} — maintain consistent character design, proportions, and colors across all 12 panels.
    
    Important Details:
    Panel 1: {scene_1} — "{caption_1}"
    Panel 2: {scene_2} — "{caption_2}"
    Panel 3: {scene_3} — "{caption_3}"
    Panel 4: {scene_4} — "{caption_4}"
    Panel 5: {scene_5} — "{caption_5}"
    Panel 6: {scene_6} — "{caption_6}"
    Panel 7: {scene_7} — "{caption_7}"
    Panel 8: {scene_8} — "{caption_8}"
    Panel 9: {scene_9} — "{caption_9}"
    Panel 10: {scene_10} — "{caption_10}"
    Panel 11: {scene_11} — "{caption_11}"
    Panel 12: {scene_12} — "{caption_12}"
    
    Panels read left-to-right, top-to-bottom (like a comic page). Character appearance, clothing, and color palette must stay identical across all 12 panels — only pose, expression, angle, and environment change.
    
    Style: {art_style}
    
    Use Case: animation pre-production, pitch deck visualization, narrative overview poster
    Constraints: all 12 panels must be distinct scenes (no duplicates), character consistency is critical, scene numbers must be legible, {language} captions
    Quality: high
    ```
    
    **Variables:**
    - `{title}` — project/episode title displayed at top
    - `{character_description}` — detailed character design (colors, outfit, distinguishing features)
    - `{scene_1}` through `{scene_12}` — visual description of each panel
    - `{caption_1}` through `{caption_12}` — text label under each panel
    - `{art_style}` — (e.g., "Pixar-style 3D", "anime cel shading", "watercolor illustration", "graphic novel ink")
    - `{language}` — caption language
    
    **Recommended size:** 1024x1536 (portrait — 3 columns x 4 rows needs vertical space)
    **Model:** GPT Image 2.5 `quality: high` (text-heavy — scene numbers and captions must be readable)
    **Common pitfalls:**
    - Character drift is the biggest risk at 12 panels — repeat character design details in the prompt, not just "same character"
    - Without explicit scene numbering in the prompt, panels may appear in random order
    - 12 panels with captions is extremely text-dense — keep captions under 5 words each for legibility
    - Art style must be stated once and applied uniformly; mixing styles across panels produces visual chaos
    
    ---
    
    ## General Multi-Panel Tips
    
    **Panel count vs. detail trade-off:** More panels = less detail per panel. 4 panels allow rich per-panel descriptions; 12 panels need brief, visually distinct descriptions (3-8 words each).
    
    **Subject consistency:** Always include an explicit instruction: "same person / same character / same product across all panels." Repeat key identity markers (hair color, clothing, distinguishing features) rather than saying "same as before."
    
    **Grid specification:** Always state the grid dimensions (e.g., "3x2 grid, 3 columns 2 rows"). Saying "6 panels" without layout instruction lets the model choose an unpredictable arrangement.
    
    **Borders and gaps:** Be explicit — "thin white border between panels" or "borderless, no gaps." Default behavior varies by model and is unreliable.
    
    **Reading order:** State it: "left-to-right, top-to-bottom" or "numbered 1-9 starting top-left." Without this, narrative flow may be jumbled.
    
    **Panel variety:** State it as a requirement, not a hope: "no two non-empty panels may be identical or near-identical — each must change the shot size, the camera angle, the action beat, or the foreground/background relationship." Without this line the model repeats a flattering composition across half the grid. It is the mirror image of subject consistency: identity stays fixed, everything else has to move.
    
    **Model selection summary:**
    - Text/labels in panels --> GPT Image 2.5 `quality: high`
    - No text, complex composition --> Nano Banana Pro
    - Budget/exploration --> Nano Banana 2 or GPT Image 2.5 `quality: low`
    
    ---
    
    ## Storyboard Sheets From a Timed Scene
    
    When the sheet feeds a video pipeline (see the sibling `video` skill), the scene's duration — not taste — decides the grid.
    
    **Grid from scene duration:**
    
    | Scene duration | Grid | Panels |
    |---|---|---:|
    | up to 9s | 3 columns x 3 rows | 9 |
    | 9-12s | 3 columns x 4 rows | 12 |
    | over 12s | 4 columns x 4 rows | 16 |
    
    Each panel stays 16:9 regardless of the grid. Keep the panels edge to edge — a thin dark rule is fine for readability, gutters are not.
    
    The duration in this table is the scene's screen time, not a clip length: the sheet shows the whole scene in one image, while the sibling `video` skill still generates it as several clips of about five seconds each. A 15s scene is one 4x4 sheet and three generations, not one.
    
    **Panel labels:** ask for two small solid-dark tabs inside every panel — the panel number in white in the top-left corner, the panel's own duration in white in the top-right ("0.5s", "1s", "2s"). Putting the duration on the panel instead of under it survives cropping and lets an editor read the sheet without the accompanying document. This is text — budget `quality: high` on GPT Image 2.5.
    
    **Default timing layouts:** start from an even split, then spend the extra time on the beat that needs a hold.
    
    - 5s on a 3x3 — eight panels at 0.5s, the last at 1s.
    - 9s on a 3x3 — nine panels at 1s.
    - 12s on a 3x4 — twelve panels at 1s.
    - 15s on a 4x4 — fourteen panels at 1s, the final two at 0.5s.
    
    Panel durations must add up to the scene duration exactly. If a short scene would fragment into meaningless beats, leave the trailing panels plain black rather than padding the action.
    
    ---
    
    *Author: Serge Shima ([t.me/aimastersme](https://t.me/aimastersme) · [sergeshima.com](https://sergeshima.com) · [aimasters.me](https://aimasters.me)) · License: CC BY 4.0 — attribution required · Source: [smixs/visual-skills](https://github.com/smixs/visual-skills)*
    
  • image/references/nano-banana.md8.8 KB66b53fae2bd6upstream ↗
    # Nano Banana — Specific Rules
    
    Три тира. «Thinking» модели — понимают намерение, физику, композицию.
    
    | Тир | Model ID | Роль |
    |-----|----------|------|
    | **NB2 Lite** | `gemini-3.1-flash-lite-image` | Самая быстрая и дешёвая ($0.034, ~4 сек), только 1K. Черновики, массовые батчи |
    | **NB2** (Flash) | `gemini-3.1-flash-image` | Дефолтный workhorse: Pro-фичи (текст, 4K, grounding, консистентность) на скорости Flash |
    | **NBP** (Pro) | `gemini-3-pro-image` | Топ-тир: сложные многослойные сцены, максимум reasoning и контроля |
    
    ## Стиль промпта
    
    Натуральный язык, 1-2 параграфа. Структура свободная, но порядок помогает:
    **Subject + Action + Location/context + Composition + Style**
    
    ```
    A cinematic wide shot of a futuristic sports car speeding through a rainy
    Tokyo street at night. Neon signs reflect off wet pavement and metallic
    chassis. Format: 16:9.
    ```
    
    ## Что НЕ указывать
    
    - Числовые параметры объектива: **50mm, 85mm, f/2.8, ISO 400** — NB игнорит. Используй описание: «shallow depth of field», «wide-angle distortion».
    - Tag-soup: «cool, modern, 4k, cinematic» — пишет связным предложением.
    
    ## Уникальные возможности
    
    ### Image Grounding (NB2 only)
    NB2 ищет реальные изображения в интернете до генерации. Архитектурно точные конкретные локации, корректные виды животных/растений.
    
    ```
    Generate a cinematic, golden-hour photograph of [SPECIFIC REAL PLACE].
    Ensure the architectural details, the spire, the surrounding square, and
    the landscape are accurate to reality.
    ```
    
    **Работает:** здания, мосты, площади, виды животных, виды растений, насекомые.
    **Не работает:** конкретные люди.
    
    ### Extreme Aspect Ratios (NB2 only)
    
    1:8, 8:1, 1:4, 4:1 — для баннеров, скроллов, комикс-стрипов.
    
    ```
    Create a 4-panel horizontal comic strip (aspect ratio 4:1).
    The story follows [CHARACTER] doing [ACTION] that ends with a twist.
    Use a vibrant comic book style. Keep character design consistent.
    ```
    
    Стандартные: 1:1, 3:2, 2:3, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9.
    
    ### Thinking Mode
    
    Включён по умолчанию; у NBP отключить нельзя (модель рисует до 2 «thought images» в бэкенде, они не тарифицируются как картинки, но thinking-токены платные). У NB2 есть рычаг `thinking_level: minimal | high` (default `minimal`) — поднимай до `high` для:
    - Невнятных результатов, требующих reasoning
    - Сложных инфографик со spatial logic
    - Grounding + spatial reasoning одновременно
    
    ### Multi-Reference (лимиты по тирам)
    
    | | NB2 Lite | NB2 | NBP |
    |---|---|---|---|
    | High-fidelity объекты | до 14 | до 10 | до 6 |
    | Character-consistency рефы | нет | до 4 | до 5 |
    | Style рефы | нет | нет | до 3 |
    
    ```
    [Image 1: face]
    [Image 2: outfit]
    [Image 3: background]
    Combine: face from Image 1, outfit style from Image 2, setting from Image 3.
    Match lighting and perspective.
    ```
    
    ### JSON для сложных сцен (5+ элементов)
    
    ```json
    {
      "subject": {"description": "main subject", "expression": "emotion/pose"},
      "photography": {"angle": "eye-level", "shot_type": "waist-up", "aspect_ratio": "16:9"},
      "background": {"setting": "location", "lighting": "soft natural"}
    }
    ```
    
    ## Editing у Nano Banana
    
    Conversational, без масок:
    
    ```
    Remove [OBJECT] from this image.
    Fill with [LOGICAL REPLACEMENT] matching surroundings.
    Keep [PRESERVED ELEMENTS] exactly the same.
    ```
    
    После генерации добавляй итеративно:
    - «Make it warmer»
    - «Increase contrast»
    - «Soften background, add blur»
    - «Change headline color to #3b82f6»
    
    Interactions API держит контекст сессии — до 3 последовательных правок стэкаются без потери исходника. Правило «Edit, don't re-roll»: картинка верна на 80% — проси точечное изменение, а не новую генерацию.
    
    ## Text Rendering
    
    - SOTA по 100+ языкам.
    - Font можно называть: «Century Gothic 12px», «Brush Script», «Impact», «Heavy blocky sans-serif».
    - Multi-language в одном кадре работает.
    - Hack: для сложного текста — сначала попроси модель **написать текст**, потом отдельным промптом «вставь этот текст в картинку».
    
    ## Resolution / Cost
    
    | Resolution | Использование |
    |------------|---------------|
    | 0.5K | Самая дешёвая, batch и А/B варианты |
    | 1K | Default |
    | 2K | Финальный отбор |
    | 4K | Печать, hero-ассеты |
    
    **Workflow:** прогон вариантов на `0.5K` flash (или NB2 Lite на 1K) → отбор → переген победителя на `2K`/`4K`.
    
    В API `image_size` пишется строго с заглавной K (`1K`, `2K`, `4K`) — `1k` отклоняется. `512px` есть только у NB2; Lite — только 1K. Без указания размера модель матчит вход, иначе 1:1.
    
    ## Кейфреймы для видео
    
    Nano Banana — image-слой перед motion-слоем. Всё критичное фиксируется на картинке ДО анимации: identity, композиция, aspect ratio, видимый текст, состояние продукта. Видеомодель усиливает любую неоднозначность стилла.
    
    - **Character sheet / hero frame.** Утверди один кадр-эталон персонажа, на него ссылаются все последующие video-промпты.
    - **Aspect ratio сразу целевой.** Кампании нужен вертикальный ролик — генери 9:16, а не кропай после анимации (тихая точка отказа).
    - **Хэндофф в видеомодель — motion brief, не пересказ.** Движение камеры, движение субъекта, движение фона, длительность, framing lock, запрещённые изменения. Видео-промпт не переописывает то, что уже есть в кадре.
    - Downstream любой: Veo 3.1, Kling 3.0, Seedance 2.0/2.5, Gemini Omni Flash (штатная пара Google, 10-сек клипы). Батч картинок → ревью-гейт → в видео уходят только одобренные стиллы.
    
    ## Известные фейлы
    
    - Руки и лица всё ещё плывут (пальцы, суставы, дрейф likeness) — явные anatomy-инструкции + рефы + перегенерации. Есть отчёты, что NBP держит лица хуже NB2.
    - Мелкий текст мылится на 1K — для плотного текста 2K+ или GPT Image 2.
    - Инфографика может содержать фактически неверные данные — цифры проверять всегда, grounding помогает, но не гарантирует.
    - Склонность к overcooked HDR / пересату — проси «natural contrast, no HDR look».
    - SynthID-водяной знак стоит на всех генерациях (невидимый); на Vertex дополнительно C2PA.
    
    ## Когда переключаться NB2 → NBP
    
    - NB2 не справляется со сложным многослойным промптом.
    - Нужен максимум референсов с разбивкой по ролям (6 объектов + 5 character + 3 style).
    - Photoreal с очень тонкой работой по материалам и свету.
    - NBP медленнее и дороже; на лицах иногда проигрывает NB2 — для портретной серии сначала тестируй оба.
    
    ---
    
    *Author: Serge Shima ([t.me/aimastersme](https://t.me/aimastersme) · [sergeshima.com](https://sergeshima.com) · [aimasters.me](https://aimasters.me)) · License: CC BY 4.0 — attribution required · Source: [smixs/visual-skills](https://github.com/smixs/visual-skills)*
    
  • image/references/patterns/character-design.md12 KB632efe8aec22upstream ↗
    # Character Design Patterns
    
    Reusable prompt templates for character turnarounds, expression sheets, outfit variants, and collectible/card formats. Each pattern uses `{variables}` for customization. Default model: GPT Image 2.5 (5-slot format) unless noted.
    
    ---
    
    ## Character Turnaround Sheet (3-View)
    
    Use for game or animation pre-production — front, side, and back views of a character on a single white canvas with color callouts and height reference lines.
    
    <!-- Source concept: character model sheet / turnaround sheet for 3D modelers and animators -->
    
    ```
    Scene: clean white background, flat even lighting with no cast shadows, thin horizontal height reference lines in #CCCCCC spanning the full width at head, shoulder, waist, knee, and foot level
    Subject: {character_name} — {character_description} — shown in three views arranged left to right: front view (facing camera), three-quarter side view (turned 75 degrees right), back view (facing away)
    Important Details: all three views share identical proportions and vertical alignment along the height lines, feet on the same baseline, arms relaxed at sides for clear silhouette reading, color callout annotations — small circles of each key color ({color_1} {hex_1}, {color_2} {hex_2}, {color_3} {hex_3}) connected by thin #999999 leader lines to the corresponding material on the front view, fine line weight for height markers, clothing folds and seam lines consistent across all three angles
    Use Case: game character pre-production, animation model sheet, 3D reference handoff
    Constraints: no background elements, no props in hands, no dramatic poses — neutral standing pose only, all three views must depict the exact same character with identical outfit and proportions, no drop shadows, no gradient backgrounds
    ```
    
    **Key levers:** `{character_name}`, `{character_description}` (age, build, hairstyle, outfit — be specific), `{color_1}`/`{hex_1}` through `{color_3}`/`{hex_3}` (key palette colors for callouts, e.g. jacket navy #1B2A4A, skin warm beige #D4A574, hair auburn #8B3A2F)
    
    **Recommended model:** GPT Image 2.5 (`quality: high`) — height reference lines and color callout text require precise rendering
    
    ---
    
    ## Expression Sheet
    
    Use to produce a grid of 6–9 facial expressions for the same character — consistent head angle and art style, with emotion labels under each face.
    
    <!-- Source concept: character expression/emotion reference sheet for animation or visual novel production -->
    
    ```
    Scene: white background, 3x3 grid (or 3x2 for 6 expressions) with thin #DDDDDD divider lines, each cell contains one head-and-shoulders portrait
    Subject: {character_name} — {character_description} — same head angle (three-quarter left), same hairstyle, same lighting across all cells
    Important Details:
      Cell 1: neutral — relaxed brow, closed mouth, calm eyes. Label: "NEUTRAL"
      Cell 2: happy — genuine smile reaching the eyes, raised cheeks. Label: "HAPPY"
      Cell 3: angry — furrowed brow, clenched jaw, narrowed eyes. Label: "ANGRY"
      Cell 4: sad — downturned mouth, glistening eyes, slightly lowered head. Label: "SAD"
      Cell 5: surprised — wide eyes, raised eyebrows, open mouth. Label: "SURPRISED"
      Cell 6: disgusted — wrinkled nose, upper lip raised, squinted eyes. Label: "DISGUSTED"
      {extra_expressions}
      Labels in small {label_font} beneath each cell, #555555 text
      Consistent {art_style} rendering across all cells — skin tone, line weight, shading approach identical
    Use Case: animation expression reference, visual novel sprite guide, game character documentation
    Constraints: same character identity in every cell — no variation in hair, accessories, skin tone, or clothing between expressions, head angle stays fixed, no hand gestures, labels must be legible
    ```
    
    **Key levers:** `{character_name}`, `{character_description}` (face shape, skin tone, hair, distinguishing marks), `{art_style}` (clean cel-shaded anime, painterly semi-realism, flat vector illustration), `{label_font}` (condensed sans-serif, monospace, rounded sans), `{extra_expressions}` (add cells 7-9: e.g. "Cell 7: smirk — one corner of mouth raised, knowing look. Label: 'SMIRK'")
    
    **Recommended model:** GPT Image 2.5 (`quality: high`) — text labels and consistent facial identity across 9 cells need precise control
    
    ---
    
    ## Outfit Variant Grid
    
    Use to show one character in multiple outfits or costumes — for fashion exploration, game skin concepts, or wardrobe design.
    
    <!-- Source concept: character costume/skin variant sheet for fashion or game design -->
    
    ```
    Scene: light neutral background ({bg_color}), {grid_layout} grid, thin white gaps between cells, soft even front lighting in every cell
    Subject: {character_name} — {character_description} — same pose ({pose_description}) in every cell, only the outfit changes
    Important Details:
      Cell 1: {outfit_1_name} — {outfit_1_description}
      Cell 2: {outfit_2_name} — {outfit_2_description}
      Cell 3: {outfit_3_name} — {outfit_3_description}
      Cell 4: {outfit_4_name} — {outfit_4_description}
      Cell 5: {outfit_5_name} — {outfit_5_description}
      Cell 6: {outfit_6_name} — {outfit_6_description}
      Outfit name in small bold sans-serif centered below each cell, #333333 text
      Character body proportions, face, hairstyle, and skin tone identical across all cells — only clothing and accessories differ
    Use Case: game skin lineup, fashion mood board, costume design exploration
    Constraints: same character identity and pose in every cell, no background scenery — character only, outfit labels must be readable, no overlapping garments between cells, {grid_layout} layout must be uniform
    ```
    
    **Key levers:** `{character_name}`, `{character_description}` (build, face, hair — anchor identity), `{bg_color}` (#F0F0F0 light gray, #FFF8F0 warm cream, #E8EDF2 cool blue-gray), `{grid_layout}` (2x3 or 3x3), `{pose_description}` (hands on hips, relaxed standing, one hand raised), `{outfit_N_name}` / `{outfit_N_description}` (e.g. "Street Casual" — oversized denim jacket, white tee, black cargo pants, chunky sneakers)
    
    **Recommended model:** GPT Image 2.5 (`quality: medium`) — character consistency is the priority; `high` only if outfit labels need fine legibility
    
    ---
    
    ## Chibi / Mini-Me Collectible
    
    Use to transform a realistic character into a cute 3D collectible figurine — oversized head, compact body, multiple poses performing different activities.
    
    <!-- Source concept: chibi/super-deformed vinyl collectible figurine with consistent identity across poses -->
    
    ```
    Scene: soft gradient background ({bg_gradient}), studio product lighting with rim light from behind and soft fill from front, subtle ground shadow beneath each figurine
    Subject: chibi-style 3D collectible figurine of {character_name} — {character_description_simplified} — large head (roughly 1:2 head-to-body ratio), rounded limbs, smooth matte vinyl surface, {num_poses} poses arranged in a row
    Important Details:
      Pose 1: {pose_1_description}
      Pose 2: {pose_2_description}
      Pose 3: {pose_3_description}
      {extra_poses}
      Face retains recognizable features from the original character — {face_markers} — simplified into the chibi style with large round eyes and small nose/mouth
      Outfit matches the original: {outfit_simplified} — colors preserved ({color_palette}), details reduced to clean shapes
      Each figurine sits on a small circular base ({base_color} matte finish)
      Vinyl toy aesthetic — visible seam line at sides, slight glossy highlight on forehead and cheeks
    Use Case: merchandise concept, social media avatar set, fan collectible design
    Constraints: consistent face and outfit across all poses, chibi proportions must stay uniform (no realistic proportions creeping in), no text on bases, smooth render — not cel-shaded, matte vinyl material throughout
    ```
    
    **Key levers:** `{character_name}`, `{character_description_simplified}` (key outfit and hair only), `{face_markers}` (e.g. round glasses, scar on left cheek, green eyes), `{color_palette}` (hex values for 2-3 dominant colors), `{bg_gradient}` (#F5F0EB to #FFFFFF warm, #E0E8F0 to #FFFFFF cool), `{num_poses}` (3-5), `{pose_N_description}` (e.g. sitting cross-legged reading a book, waving with both hands, holding a coffee cup), `{base_color}` (white, black, matching character's main color)
    
    **Recommended model:** GPT Image 2.5 (`quality: medium`) — smooth 3D vinyl surfaces render well at medium; `high` for marketing-ready close-ups
    
    ### Nano Banana version:
    
    ```
    A row of {num_poses} chibi-style 3D vinyl collectible figurines of {character_name}, each in a different pose. Large head (1:2 head-to-body ratio), rounded limbs, smooth matte vinyl surface with visible seam lines and subtle glossy highlights on the forehead. Face retains {face_markers} simplified into chibi proportions with large round eyes. Outfit: {outfit_simplified} in {color_palette}. Poses left to right: {pose_1_description}, {pose_2_description}, {pose_3_description}. Each figurine on a small circular {base_color} matte base. Soft gradient background ({bg_gradient}), studio product lighting with rim light from behind and soft fill from front, subtle ground shadow. Format: 16:9.
    ```
    
    ---
    
    ## Anime-Style Character Card
    
    Use for a full character reference card with portrait, full body, key items, and color palette — organized on a white background in a professional concept art layout.
    
    <!-- Source concept: anime/game character reference sheet with stats, items, and palette swatches -->
    
    ```
    Scene: white background, organized reference card layout divided into clear sections with thin #CCCCCC separator lines
    Subject: {character_name} — {character_description} — anime-style rendering with clean line art and flat cel shading
    Important Details:
      Left section (40% width): full-body standing pose, front-facing, arms slightly away from body to show full outfit, feet visible, confident neutral expression
      Upper-right section: portrait bust — head and shoulders, three-quarter angle, detailed face rendering showing {face_details}
      Mid-right section: {num_items} key items arranged in a row — {item_1}, {item_2}, {item_3} — each drawn at consistent scale with thin outline, labeled in small text below
      Lower-right section: color palette — {num_swatches} rectangular swatches in a horizontal strip showing the character's key colors ({swatch_colors}), hex code below each swatch
      Bottom strip: brief stat block or bio text in clean sans-serif — "Name: {character_name} | Class: {class} | Height: {height} | Affiliation: {affiliation}"
      Consistent line weight and shading style across all sections
    Use Case: game character documentation, light novel illustration guide, animation production reference
    Constraints: unified anime art style across all sections — portrait and full body must be the same character with identical design, items must match what the character wears/carries in the full body view, no decorative borders or ornamental frames, text must be legible at screen resolution
    ```
    
    **Key levers:** `{character_name}`, `{character_description}` (detailed: hair color/style, eye color, outfit layers, accessories), `{face_details}` (distinctive facial features — e.g. heterochromia, facial tattoo, sharp jawline), `{item_1}`/`{item_2}`/`{item_3}` (signature weapon, accessory, artifact), `{swatch_colors}` (e.g. midnight blue #191970, cherry red #C41E3A, silver #C0C0C0, warm skin #E8B89D), `{num_swatches}` (4-6), `{class}` / `{height}` / `{affiliation}` (stat block fields)
    
    **Recommended model:** GPT Image 2.5 (`quality: high`) — text-heavy layout with hex codes, labels, and stat block requires precise rendering
    
    ### Nano Banana version:
    
    ```
    An anime-style character reference card for {character_name} on a white background. Left side: full-body standing pose, front-facing, clean cel-shaded anime rendering — {character_description}. Upper right: portrait bust at three-quarter angle showing {face_details}. Mid-right: key items laid out in a row — {item_1}, {item_2}, {item_3} — each drawn consistently and labeled. Lower right: horizontal color palette strip with {num_swatches} rectangular swatches ({swatch_colors}). Clean layout with thin gray separator lines between sections. Professional concept art quality, consistent line weight throughout. Format: 3:4.
    ```
    
    ---
    
    *Author: Serge Shima ([t.me/aimastersme](https://t.me/aimastersme) · [sergeshima.com](https://sergeshima.com) · [aimasters.me](https://aimasters.me)) · License: CC BY 4.0 — attribution required · Source: [smixs/visual-skills](https://github.com/smixs/visual-skills)*
    
  • image/references/patterns/ecommerce.md9.8 KB4020e34585ddupstream ↗
    # E-Commerce Product Photography Patterns
    
    Reusable prompt templates for product ads, packaging, and commercial visuals. Each pattern uses `{variables}` for customization. Default model: GPT Image 2.5 (5-slot format) unless noted.
    
    ---
    
    ## Miniature Diorama Product Ad
    
    Use when you need a playful, attention-grabbing product visual where tiny workers interact with an oversized product — ideal for social media ads and launch campaigns.
    
    <!-- Source concept: miniature/tilt-shift product advertising with construction-worker scale play -->
    
    ```
    Scene: clean {surface_color} tabletop studio, soft overhead diffused light, shallow depth of field with tilt-shift blur at edges
    Subject: oversized {product_name} centered in frame, surrounded by miniature construction workers (1/87 scale figurines) — some climbing the packaging with tiny ladders, others operating a miniature crane to lift the cap/lid, a crew painting the label
    Important Details: {product_material} surface catches key light from upper left, visible condensation/texture on product, figurines cast tiny hard shadows, warm {accent_color} hard hats on workers, fine detail on miniature tools, slightly desaturated background, eye-level camera angle
    Use Case: social media product ad, Instagram carousel hero
    Constraints: no text, no logos other than product label, no floating elements, product must remain recognizable and unaltered, no more than 8 figurines
    ```
    
    **Key levers:** `{product_name}`, `{product_material}` (frosted glass, matte aluminum, glossy plastic), `{surface_color}` (white marble, raw concrete, light birch wood), `{accent_color}` (safety orange #FF6600, yellow #FFD600)
    
    **Recommended model:** GPT Image 2.5 (`quality: high`) — precise figurine detail and product label legibility
    
    ---
    
    ## Luxury Cosmetics Studio Shot
    
    Use for premium beauty or fragrance product photography — dark, moody, tactile surfaces with atmospheric effects.
    
    <!-- Source concept: luxury perfume/cosmetics dark-marble studio photography with condensation and smoke -->
    
    ```
    Scene: dark studio, {background_surface} surface with subtle reflections, thin layer of low-hanging smoke drifting left to right, {time_mood} ambient
    Subject: {product_name} bottle/tube centered, three-quarter angle, {product_finish} catching a single key light from upper right
    Important Details: fine water droplets on product surface (condensation, not spray), {accent_material} accent elements flanking the product (raw stone, dried botanicals, metal shavings), volumetric haze behind product, reflection on surface below is soft and dark, color palette restricted to {palette}, contact shadow sharp near base fading to soft
    Use Case: luxury brand campaign hero, print ad, website banner
    Constraints: no text overlays, no human hands, background stays dark (#0a0a0a to #1a1a1a gradient), no color spill outside the defined palette, no lens flare
    ```
    
    **Key levers:** `{background_surface}` (nero marquina marble, wet obsidian slab, brushed gunmetal), `{product_finish}` (frosted glass, lacquered black, brushed gold), `{accent_material}` (raw quartz crystals, dried lavender stems, black river stones), `{palette}` (golds #C9A84C and blacks, rose #B76E79 and creams, emerald #2D6A4F and silvers), `{time_mood}` (cold blue, warm amber)
    
    **Recommended model:** GPT Image 2.5 (`quality: high`) — surface materials and condensation detail
    
    ---
    
    ## 9-Panel TVC Storyboard Grid
    
    > For a dark-themed variant with timestamps, see [multi-panel.md](../multi-panel.md#1-9-cell-grid-storyboard).
    
    Use to present a product commercial shot breakdown in a single image — pitch decks, creative presentations, client approvals.
    
    <!-- Source concept: 9-panel television commercial storyboard grid with numbered frames -->
    
    ```
    Scene: white canvas background, clean 3x3 grid with thin #CCCCCC divider lines (2px), each cell represents one shot in a {duration}-second commercial
    Subject: {product_name} commercial storyboard — each panel is a distinct camera setup
    Important Details:
      Panel 1 (wide): establishing shot of {setting}, warm natural light, product not yet visible
      Panel 2 (medium): {protagonist} notices/discovers the product on {surface}
      Panel 3 (close-up): hand reaching for {product_name}, shallow depth of field
      Panel 4 (ECU): product detail — texture of {product_material}, label readable
      Panel 5 (medium): {protagonist} using/opening the product, genuine expression
      Panel 6 (reaction): close-up face, {emotion} expression, soft key light
      Panel 7 (wide): product in context of {lifestyle_scene}
      Panel 8 (beauty shot): product hero on {beauty_surface}, studio lighting
      Panel 9 (pack shot): product centered on white with "{tagline}" below in thin sans-serif, #333333
    Use Case: creative pitch deck, storyboard for TVC production
    Constraints: consistent character identity across all panels, no panel numbering text, uniform lighting temperature within the narrative (panels 1-7), distinct studio lighting for panels 8-9
    ```
    
    **Key levers:** `{product_name}`, `{protagonist}` (woman in her 30s, young couple, family), `{setting}` (bright kitchen, outdoor terrace, urban cafe), `{emotion}` (satisfied, surprised, relaxed), `{beauty_surface}` (white marble, gradient gray), `{tagline}`, `{duration}` (15, 30)
    
    **Recommended model:** GPT Image 2.5 (`quality: high`) — grid precision and text in panel 9
    
    ---
    
    ## Floating Ingredient Freeze-Frame
    
    Use for food, beverage, or supplement products where suspended ingredients communicate freshness, flavor, or composition.
    
    <!-- Source concept: frozen-motion ingredient explosion around product, high-speed photography aesthetic -->
    
    ```
    Scene: {background_gradient} gradient backdrop, high-speed flash freeze-frame moment, clean studio environment
    Subject: {product_name} container in center, tilted {tilt_angle} degrees, with {liquid_type} mid-pour arcing from the opening
    Important Details: individual {ingredient_list} frozen in mid-air around the product — each element sharply focused with visible texture ({texture_details}), micro water droplets suspended alongside ingredients, single hard flash from behind (rim light on ingredients), secondary soft fill from front, liquid splash forms a clean arc with visible viscosity, product label faces camera and remains fully legible, ingredients distributed in a loose orbital pattern
    Use Case: beverage packaging, food product poster, social media ad
    Constraints: no ingredients overlapping the product label, no motion blur (everything frozen sharp), background must remain clean — no stray splashes hitting edges, no more than {max_ingredients} floating elements, no artificial glow effects
    ```
    
    **Key levers:** `{product_name}`, `{background_gradient}` (#F5F0EB to #FFFFFF for light, #1A0A2E to #0D0D0D for dark), `{liquid_type}` (amber juice, white milk, green smoothie), `{ingredient_list}` (sliced strawberries + mint leaves + ice cubes, cocoa nibs + hazelnuts + vanilla pod), `{texture_details}` (visible seeds on strawberry cross-section, frost crystals on ice), `{tilt_angle}` (15, 25), `{max_ingredients}` (6-8)
    
    **Recommended model:** GPT Image 2.5 (`quality: high`) — frozen detail precision and label legibility
    
    ---
    
    ## Inflatable Surrealism Product Poster
    
    Use for disruptive, scroll-stopping social ads where the product packaging appears squeezed, inflated, or physically distorted as if made of soft rubber or vinyl.
    
    <!-- Source concept: inflatable surrealism — product packaging rendered as squeezed/puffy/distorted soft objects -->
    
    ```
    Scene: solid {background_color} background, soft even studio lighting with no hard shadows, slightly elevated camera angle (15 degrees above eye level)
    Subject: {product_name} packaging reimagined as a puffy inflatable vinyl object — the shape is recognizable but squeezed at the middle as if gripped by an invisible hand, seams visible where vinyl panels meet, surface slightly reflective like a pool float
    Important Details: {product_color_scheme} preserved on the inflated surface but stretched and slightly warped around curves, visible air valve at the bottom edge (small brass circle), subtle wrinkles where the vinyl compresses, the brand name/logo distorted by the inflation but still readable, two or three {companion_objects} nearby also inflated (matching aesthetic), environment reflection on glossy vinyl surface, cast shadow soft and diffused below
    Use Case: disruptive social media ad, brand campaign poster, billboard
    Constraints: product must remain identifiable despite distortion, no liquid, no particles, no humans, no text outside what exists on the packaging, vinyl texture must read as physical (not digital 3D render), background is flat color only
    ```
    
    **Key levers:** `{product_name}`, `{background_color}` (bubblegum pink #FFB6C1, electric blue #007BFF, acid yellow #E8FF00), `{product_color_scheme}`, `{companion_objects}` (matching accessories, ingredient items, brand mascot elements)
    
    **Recommended model:** NBP — complex spatial reasoning for believable physical distortion
    
    ### Nano Banana version:
    
    ```
    A product poster showing {product_name} packaging transformed into a puffy inflatable vinyl object, squeezed at the middle as if gripped by an invisible hand. The surface is slightly glossy like a pool float, with visible seams where vinyl panels meet and a small brass air valve at the base. The original {product_color_scheme} is preserved but stretched and warped around the inflated curves. Brand text is distorted by the shape but still legible. Two small {companion_objects} sit nearby, also inflated in the same vinyl style. Solid {background_color} background, soft even studio light, slightly elevated camera angle. Soft diffused shadow below. Format: 4:5.
    ```
    
    ---
    
    *Author: Serge Shima ([t.me/aimastersme](https://t.me/aimastersme) · [sergeshima.com](https://sergeshima.com) · [aimasters.me](https://aimasters.me)) · License: CC BY 4.0 — attribution required · Source: [smixs/visual-skills](https://github.com/smixs/visual-skills)*
    
  • image/references/patterns/fashion-editorial.md11 KBee0a0a4fd117upstream ↗
    # Fashion Editorial Patterns
    
    Reusable prompt templates for fashion campaigns, lookbooks, and editorial shoots. Each pattern uses `{variables}` for customization. Default model: GPT Image 2.5 (5-slot format) unless noted.
    
    ---
    
    ## 3-Panel Campaign Collage
    
    Use for fashion brand campaign hero images — one wide shot combining hero pose, close-up detail, and action/movement in a triptych layout.
    
    <!-- Source concept: fashion campaign triptych collage — hero + detail + movement panels -->
    
    ```
    Scene: three vertical panels side by side on a single canvas (left 40% width, right split into two stacked panels 60% height each), thin white divider lines (3px), cohesive {lighting_setup} across all panels
    Subject: {model_description} wearing {outfit_description} — same person, same outfit, three perspectives
    Important Details:
      Left panel (hero): full-body three-quarter pose, {model_description} standing against {background_1}, weight shifted to one hip, direct gaze at camera, medium-format film grain, {color_grade}
      Top-right panel (detail): extreme close-up of {detail_focus} — visible weave/stitching/texture of {fabric_type}, shallow depth of field, warm directional light raking across surface
      Bottom-right panel (action): mid-stride walking shot from low angle, {background_2}, coat/fabric in motion with natural drape physics, slight motion energy in hair
      Overall: cohesive warm/cool temperature of {color_temperature}, consistent skin tone rendering across panels, editorial magazine quality
    Use Case: fashion brand campaign, lookbook cover, social media carousel hero
    Constraints: same person with identical features in all three panels, no text, no logos, no visible studio equipment, panels must feel like one shoot not three separate photos, no heavy retouching glow
    ```
    
    **Key levers:** `{model_description}` (East Asian woman mid-20s, athletic man early 30s), `{outfit_description}`, `{detail_focus}` (collar construction, cuff button, belt buckle, shoe sole), `{fabric_type}` (raw selvedge denim, double-faced cashmere, washed silk), `{background_1}` (concrete wall, sand dune, industrial corridor), `{background_2}` (open street, field, rooftop), `{color_grade}` (lifted blacks with amber cast, desaturated teal), `{color_temperature}` (warm 4000K feel, cool overcast daylight)
    
    **Recommended model:** GPT Image 2.5 (`quality: high`) — identity consistency across panels and fabric texture detail
    
    ---
    
    ## 2x2 Editorial Portrait Grid
    
    Use for model tests, casting cards, or editorial portfolio pages — four angles of the same person in a clean grid.
    
    <!-- Source concept: 2x2 fashion portrait grid — same model, four setups -->
    
    ```
    Scene: 2x2 grid on white canvas, thin {divider_color} divider lines (2px), all four cells use the same {background_type} with slight variation in angle
    Subject: {model_description}, same {outfit_description} in all four frames
    Important Details:
      Top-left: straight-on headshot, neutral expression, eyes to camera, even butterfly lighting
      Top-right: three-quarter profile, chin slightly lifted, single key light from camera-left creating defined cheekbone shadow
      Bottom-left: full profile silhouette, rim light from behind outlining jaw and nose, {background_type} slightly darker
      Bottom-right: candid moment — mid-laugh or adjusting {accessory}, natural movement, softer light
      All frames: {film_stock} color science, consistent {skin_tone_handling}, no heavy skin smoothing — visible pores and natural texture, shallow depth of field in all four
    Use Case: fashion editorial spread, model comp card, casting portfolio
    Constraints: identical person in all four frames (bone structure, skin, hair must match exactly), no makeup changes between frames, no text, no watermarks, backgrounds must feel cohesive not random
    ```
    
    **Key levers:** `{model_description}`, `{outfit_description}` (black turtleneck, white linen shirt unbuttoned at collar), `{background_type}` (seamless medium gray, textured plaster wall, out-of-focus greenery), `{divider_color}` (#FFFFFF, #E0E0E0), `{film_stock}` (Kodak Portra 400, Fujifilm Pro 400H), `{skin_tone_handling}` (warm undertones preserved, cool-neutral rendering), `{accessory}` (earring, collar, watch)
    
    **Recommended model:** GPT Image 2.5 (`quality: high`) — identity consistency critical across four frames
    
    ---
    
    ## Streetwear Poster with Oversized Typography
    
    Use for streetwear drops, limited edition launches, or urban fashion brand campaigns where bold type dominates the composition.
    
    <!-- Source concept: streetwear poster with model integrated into oversized typographic layout -->
    
    ```
    Scene: solid {background_color} background, graphic poster composition, typography dominates 60% of visual space
    Subject: {model_description} in {streetwear_outfit}, standing or crouching in a {pose_description}, positioned {model_position} — partially overlapping the text layers
    Important Details: headline "{HEADLINE_TEXT}" in extra-bold condensed {font_style}, {text_color}, occupying upper two-thirds of frame — model breaks in front of some letters and behind others (depth interplay), secondary text "{SUBHEAD_TEXT}" in thin weight {subhead_color} near bottom edge, {lighting_type} on model creating {shadow_quality} shadows, grain overlay across entire image ({grain_intensity}), composition follows rule of thirds with model at {grid_position} intersection
    Use Case: streetwear brand drop poster, social media announcement, lookbook cover
    Constraints: text must be fully legible even where model overlaps, no extra text or watermarks, no decorative elements besides type and model, model does not obscure more than 30% of any single letter, "{HEADLINE_TEXT}" spelled exactly as provided
    Quality: high
    ```
    
    **Key levers:** `{HEADLINE_TEXT}` (brand name, drop name), `{SUBHEAD_TEXT}` (date, "LIMITED DROP", collection name), `{background_color}` (off-white #F5F1EB, concrete gray #8C8C8C, matte black #0D0D0D), `{text_color}` (#000000, #FF3333, #FFFFFF), `{font_style}` (sans-serif like Druk Wide, slab-serif, stencil), `{streetwear_outfit}`, `{model_position}` (center-left, right third), `{lighting_type}` (harsh direct flash, soft window light), `{grain_intensity}` (subtle film grain, heavy 35mm grain)
    
    **Recommended model:** GPT Image 2.5 (`quality: high`) — text rendering and model-type depth interplay
    
    ---
    
    ## Retro Roller Skating Sportswear Campaign
    
    Use for playful, nostalgic sportswear or athleisure campaigns with 70s-80s visual language.
    
    <!-- Source concept: retro roller skating / sportswear campaign with analog film aesthetic -->
    
    A sun-drenched wide shot of {model_description} roller skating along a {location_description}. They wear {outfit_description} — the fabric catches light as they move, one leg extended mid-glide, arms relaxed and swinging naturally. The ground is smooth asphalt with painted lane markings in faded {lane_color}. Background shows {background_elements} slightly out of focus through heat haze. Shot on {film_stock} with pronounced grain and slightly lifted shadows. Color palette centers on {palette_description}. Golden hour backlight creates a warm halo around the subject and long shadow stretching toward camera. Genuine movement energy — hair and loose fabric respond to speed. Format: {aspect_ratio}.
    
    **Key levers:** `{model_description}`, `{outfit_description}` (high-waisted terry shorts in coral, cropped zip-up in cream, tube socks with racing stripes), `{location_description}` (Venice Beach boardwalk, empty suburban tennis court, coastal promenade), `{film_stock}` (Kodak Gold 200, Fuji Superia 400), `{palette_description}` (terracotta #CC5533, cream #FFF5E1, sky blue #87CEEB, mustard #D4A017), `{lane_color}` (faded yellow, sun-bleached white), `{background_elements}` (palm trees and pastel buildings, chain-link fence and bleachers), `{aspect_ratio}` (3:2, 16:9)
    
    **Recommended model:** NB2 — natural movement, analog film grain, atmospheric grounding
    
    ---
    
    ## Futuristic Sportswear Editorial with 3D Blob Shapes
    
    Use for forward-looking athletic or techwear editorials where abstract 3D forms create a surreal spatial environment around the model.
    
    <!-- Source concept: futuristic sportswear editorial with organic 3D blob/sphere shapes -->
    
    ```
    Scene: {studio_environment} studio space, matte {floor_color} floor extending to infinity, ambient fill light with no visible source, three to five large organic 3D blob shapes ({blob_color}, glossy smooth surface with environment reflections) floating at varying heights around the subject
    Subject: {model_description} in {techwear_outfit}, posed in an athletic stance — {pose_detail}
    Important Details: blobs range from basketball-sized to armchair-sized, each with smooth amoebic curves and a single specular highlight, they cast soft colored shadows onto the floor and the model's clothing, model lit by cool directional light from camera-right creating defined muscle/fabric contour, {fabric_detail} visible in the garment construction, one blob partially behind the model and one in front (spatial depth), color palette limited to {palette}, overall mood is clinical and aspirational
    Use Case: sportswear lookbook, techwear campaign, editorial magazine spread
    Constraints: blobs must look physically present (not composited), no text, no logos, no additional props, model remains the clear focal point despite the surrounding forms, blobs do not touch or intersect with the model's body, no motion blur
    ```
    
    **Key levers:** `{studio_environment}` (white void, concrete gray, deep navy), `{floor_color}` (light gray #D0D0D0, charcoal #333333), `{blob_color}` (chrome silver, translucent jade #00A86B, matte coral #FF6B6B), `{techwear_outfit}` (bonded seam track pants + compression top, oversized windbreaker + utility shorts), `{pose_detail}` (low lunge position, standing with one arm extended checking a wrist device, mid-jump), `{fabric_detail}` (visible bonded seams, reflective piping, mesh ventilation panels), `{palette}` (monochrome + single accent, earth tones + neon green #39FF14)
    
    **Recommended model:** NBP — complex spatial reasoning for blob placement and reflections
    
    ### Nano Banana version:
    
    ```
    A futuristic sportswear editorial photograph in a {studio_environment} studio space with a matte {floor_color} floor extending to infinity. {model_description} stands in an athletic {pose_detail}, wearing {techwear_outfit} with {fabric_detail}. Three to five large organic 3D blob shapes in {blob_color} with glossy smooth surfaces float at varying heights around the subject — the largest is armchair-sized, the smallest basketball-sized. Each blob has amoebic curves and casts soft colored shadows onto the floor and the model's clothes. Cool directional light from camera-right defines contour and fabric texture. One blob sits partially behind the model, one in front, creating depth. The model is the clear focal point. No text, no logos. Format: 4:5.
    ```
    
    ---
    
    *Author: Serge Shima ([t.me/aimastersme](https://t.me/aimastersme) · [sergeshima.com](https://sergeshima.com) · [aimasters.me](https://aimasters.me)) · License: CC BY 4.0 — attribution required · Source: [smixs/visual-skills](https://github.com/smixs/visual-skills)*
    
  • image/references/patterns/food-beverage.md12 KB85571cf34283upstream ↗

    Contents are not inlined in the catalog. Open the upstream file.

  • image/references/patterns/portrait-cinema.md13 KBd23168ee5716upstream ↗

    Contents are not inlined in the catalog. Open the upstream file.

  • image/references/patterns/poster-illustration.md20 KBad0ad35b1733upstream ↗

    Contents are not inlined in the catalog. Open the upstream file.

  • image/references/patterns/ui-social.md14 KB80c45673ef9eupstream ↗

    Contents are not inlined in the catalog. Open the upstream file.

  • image/references/prompt-framework.md13 KB10a8b595215dupstream ↗

    Contents are not inlined in the catalog. Open the upstream file.

  • image/references/slides.md20 KB684a43f49f03upstream ↗

    Contents are not inlined in the catalog. Open the upstream file.

  • image/references/storyboards.md2.3 KB28d0986cd725upstream ↗
    # Storyboards & Sequential Art
    
    One-shot narrative sequences. Concept art. Consistent across frames.
    
    ## Story Sequence
    
    ```
    Create [NUMBER]-part story with [NUMBER] images.
    
    Characters: [DESCRIPTION]
    Setting: [LOCATION/CONTEXT]
    Story arc: [NARRATIVE with emotional highs/lows]
    
    Identity: Characters and attire consistent throughout.
    Variation: Different angles, distances, expressions.
    Generate: One image at a time.
    Format: [ASPECT RATIO] per image.
    ```
    
    ## Commercial Storyboard
    
    ```
    Create [NUMBER]-part story for [PRODUCT] commercial.
    
    Characters: [man and woman / family / solo]
    Product: [DESCRIPTION]
    Story: [EMOTIONAL ARC]
    Ending: Elegant shot with [LOGO/PRODUCT]
    
    Identity and attire consistent throughout.
    Vary angles and distances.
    Generate one at a time.
    Format: 16:9 landscape.
    ```
    
    ## Film Storyboard
    
    ```
    Create storyboard for this scene.
    
    Shots:
    1. Establishing shot - [DESCRIPTION]
    2. Medium shot - [ACTION]
    3. Close-up - [DETAIL]
    4. POV shot - [PERSPECTIVE]
    
    Style: [B&W sketch / color / cinematic]
    Annotations: [CAMERA MOVES / NOTES]
    ```
    
    ## Concept Art Series
    
    ```
    Create concept art series for [PROJECT].
    
    Subjects: [CHARACTERS / ENVIRONMENTS / PROPS]
    Style: [AESTHETIC]
    Mood: [TONE]
    
    Consistency: Design language across all pieces.
    Generate one at a time.
    ```
    
    ## Sprite Sheets
    
    ```
    Sprite sheet of [CHARACTER] doing [ACTION].
    Grid: [3x3 / 4x4]
    Sequence: frame by frame animation
    Format: square aspect ratio
    Follow reference grid structure exactly.
    ```
    
    Tip: Extract cells to make GIF.
    
    ## Consistency Rules
    
    - "Identity must stay consistent throughout"
    - "Attire same across all images"
    - "Can be seen from different angles and distances"
    - "Expressions and poses should vary"
    - "Only one of each character per image"
    
    ## Narrative Pacing
    
    - "Story is thrilling with emotional highs and lows"
    - "Ends on [happy / dramatic / mysterious] moment"
    - "Build tension through first [N] images"
    - "Climax in image [N], resolution in final"
    
    ## Output Control
    
    - "Generate images one at a time"
    - "Make sure every image is [ASPECT RATIO]"
    - "Please generate [NUMBER] images, one at a time"
    
    ---
    
    *Author: Serge Shima ([t.me/aimastersme](https://t.me/aimastersme) · [sergeshima.com](https://sergeshima.com) · [aimasters.me](https://aimasters.me)) · License: CC BY 4.0 — attribution required · Source: [smixs/visual-skills](https://github.com/smixs/visual-skills)*
    
  • image/references/structural.md2.2 KB76785793870aupstream ↗
    # Structural Control
    
    Input images control composition. Sketch→final, wireframes, grids.
    
    ## Sketch → Final
    
    ```
    Create [ASSET TYPE] following this sketch exactly.
    
    Sketch defines: composition, element placement, text position
    Quality: [polished / professional / specified style]
    Text: Render "[EXACT TEXT]" where indicated
    ```
    
    Hand-drawn layouts become polished assets.
    
    ## Wireframe → Mockup
    
    ```
    Generate [MOCKUP TYPE] following these wireframes.
    
    Layout: Match wireframe structure precisely
    Content: [placeholder / realistic]
    Style: [UI style / brand guidelines]
    ```
    
    UI/UX designs from rough wireframes.
    
    ## Grid-Based Assets
    
    ```
    Generate [ASSET] fitting this [SIZE] grid.
    Each cell: [CONTENT]
    Consistency: Uniform style across cells
    ```
    
    Use cases:
    - Sprite sheets for games
    - LED display content
    - Tile-based graphics
    
    ## Pixel Art
    
    ```
    Generate pixel art [SUBJECT] fitting [SIZE] grid.
    Colors: high contrast
    Style: [retro / modern pixel]
    ```
    
    Tip: Extract cell colors programmatically for LED matrices.
    
    ## Layout Guidance
    
    Reference images can control:
    - **Where** elements sit
    - **What** proportions to use
    - **How** text flows
    - **Which** areas get emphasis
    
    ## Brand Identity Systems
    
    ```
    [Input: logo sketch or design]
    Create identity system using this logo.
    
    Generate one at a time:
    1. Logo refinement
    2. Business card mockup
    3. Letterhead
    4. Billboard application
    5. Product packaging
    6. Digital banner
    7. Social media avatar
    8. App icon
    9. Signage
    10. Merchandise
    
    Format: 16:9 each
    Maintain: brand consistency across all
    ```
    
    ## Composition Templates
    
    Use reference for:
    - Magazine layouts
    - Poster compositions
    - Social media templates
    - Presentation structures
    
    ```
    Follow the composition of reference image.
    Replace: [ELEMENTS] with [NEW CONTENT]
    Maintain: proportions, visual hierarchy
    ```
    
    ## Precision Control
    
    - "Position text exactly where sketch indicates"
    - "Follow grid structure exactly"
    - "Match layout precisely"
    - "Keep proportions from reference"
    
    ---
    
    *Author: Serge Shima ([t.me/aimastersme](https://t.me/aimastersme) · [sergeshima.com](https://sergeshima.com) · [aimasters.me](https://aimasters.me)) · License: CC BY 4.0 — attribution required · Source: [smixs/visual-skills](https://github.com/smixs/visual-skills)*
    
  • image/references/text-rendering.md5.7 KBb2a1c0f0c75cupstream ↗

    Contents are not inlined in the catalog. Open the upstream file.

  • image/references/vision-decomposer.md6.7 KB6e4a426e5443upstream ↗

    Contents are not inlined in the catalog. Open the upstream file.

  • README.mdupstream ↗

    Contents are not inlined in the catalog. Open the upstream file.

Plain-text copy, no login: /skill-bundles/smixs-visual-skills-image.md

// REPO STATS

667 stars
1 open issues
Last commit: 9d ago