---
name: xxg-portrait-rebuild-light
description: "Edit an existing JPG, JPEG, PNG, or WebP portrait to rebuild physically coherent light, exposure, color, capture style, clean optical skin, and frame quality without changing the person. Use for plastic-skin or AI-look removal, dirty synthetic texture, banding, edit seams, color drift, multi-round quality loss, tone or white-balance correction, camera/film/device emulation, backlight, prism, blind shadows, candlelight, neon, golden hour, rim light, rain-wet low key, studio light, or silhouettes."
---

# XXG Portrait Rebuild Light V2.3

## Objective

Treat the input as the same photograph, never as a reference for a replacement portrait. Improve illumination, exposure, color rendering, capture character, skin response, and generation quality while retaining identity, facial geometry and natural asymmetry, expression, pose, camera view, framing, and subject scale. Preserve source optics unless the user explicitly requests an optical restyle.

Build realism from three separable signals:

1. clean low-frequency skin tone and broad transitions;
2. source-driven diffuse/specular response with highlights confined to plausible light-facing areas;
3. fine, region-specific microdetail limited by face scale, focus, and illumination.

Do not manufacture realism with dirt, darkness, coarse pores, uniform grain, random color patches, or exaggerated facial lines. The result must first read as a clean photograph at normal size.

## Choose the Edit Scope

- **`texture-only`**: when the user asks only to remove plastic/AI skin or recover detail, force `L0 + E0 + T0 + G0 + D0 + A0 + Q0`. Preserve the source lighting, highlight placement, exposure, white balance, focal plane, depth of field, and background.
- **`tone-and-exposure`**: when the user asks for brightness, highlight/shadow recovery, white balance, palette, or grading without new lighting, preserve the source light direction and edit only E/T/G.
- **`relight-and-skin`**: when the user requests a lighting change or selects L/T/A, authorize the requested light response while preserving source focal plane and depth of field unless explicitly changed.
- **`capture-style`**: when the user requests a camera, film, phone, CCD, or era look, edit G/D and only the exposure behavior logically required by that capture; preserve viewpoint, crop, focal plane, and depth of field.
- **`optical-restyle`**: only when the user explicitly requests a different focal-length perspective, camera angle, aperture behavior, or depth of field. State that axis once and permit only the minimum framing/background reconstruction it requires.
- **`quality-repair`**: when the user asks to remove dirty synthetic patterns, banding, posterization, seams, halos, or other generation residue, edit only Q and preserve light, color, structure, optics, and natural texture.
- **`base-color-repair`**: when an edited result has an unintended cast, use the original root base to restore only unchanged or unauthorized color relationships; preserve deliberate recolors and requested E/T/G/A changes.
- **`multi-round-repair`**: when repeated edits accumulate tint, noise, sharpening, compression, banding, or seams, use the latest accepted edit as the target and the first root base as the stable identity/geometry/color reference.

Never let texture repair become relighting, tone correction become a new light source, device emulation become a new pose/viewpoint, or quality cleanup become global smoothing or an unauthorized regrade.

## Use the Host Image Editor

1. Inspect the source and read the host's native image-generation or image-editing skill.
2. Discover the actual callable in the tool registry. In Codex, inspect `ALL_TOOLS` and prefer the exact discovered name `image_gen__imagegen`.
3. For a local source in Codex, use the discovered tool and its real arguments:

```js
const result = await tools.image_gen__imagegen({
  referenced_image_paths: ["/absolute/path/source.png"],
  prompt: "compact English image-edit prompt"
});
generatedImage(result);
```

Never guess `tools.image_gen` or `input_image`. Correct wrong members, arguments, or `TypeError` from the registered signature and retry. In Claude, OpenClaw, or another host, use the equivalent native image-edit action explicitly exposed by that host.

For base-anchored or multi-round repair, pass exactly `[CURRENT_EDIT, ROOT_BASE]` and state: Image 1 is the edit target; Image 2 is the original root reference for declared locked axes only. Do not attach every intermediate round.

## Route the Outcome

| Observed state | Required action |
| --- | --- |
| Compatible image tool discovered | Invoke it. A small face, dense text, complex props, or edge contact lowers detail ambition but never blocks generation. |
| Correct image tool returns a real error | Report the actual error, enter `prompt-only`, and return a complete compact prompt. |
| Discovery completes with no compatible callable | Enter `invocation-handoff` and return a complete compact prompt. |
| Result is nearly unchanged, changes identity, creates artificial skin, or misses the light | State that the result did not achieve the requested improvement, enter `prompt-handoff`, and recompile from the source. |

Reading a skill, inspecting an image, creating a task, or announcing generation is not an image-tool invocation.

## Never Produce the Final Image Locally

Do not use Pillow, NumPy, OpenCV, ImageMagick, FFmpeg, `sips`, or custom raster scripts to relight, grade, retouch, sharpen, add texture, resize, crop, extend, composite, repair, or produce the delivered image. Use them only for read-only aspect-ratio, mask, and result audits. See `requirements.txt`.

## Compile the Image Prompt

Read [the V2.3 prompt compiler](references/prompt-recipes.md), [the lighting and skin recipes](references/lighting-skin-color-temperature-recipes.md), [the tone/exposure/style/device recipes](references/tone-exposure-style-device-recipes.md), and [the quality/iteration recipes](references/quality-repair-and-iteration.md). Decide internally as:

```text
Scope → Key L → Exposure E → Fill/Shadow → Skin scale S → Skin finish P → Light color T → Look G → Capture D → Background/Atmosphere A → Quality Q
```

Select exactly:

```text
one L + one E + one S + one P + one T + one G + one D + zero or one A + one Q
```

Use one key-light system. Atmosphere and skin reflections must inherit its direction, size, falloff, and color. E controls exposure, G controls palette/curve, and D controls capture response; none may invent another light. `A6` is the sole override: force E6 silhouette exposure, remove all subject fill/catchlights/internal illumination, use L/T only for the rear source and background, and suppress S/P. Q may clean the background but must never restore detail inside the silhouette.

Send four core lines, adding RENDER for a requested grade/device response and QUALITY for an observed repair target:

```text
EDIT: scope, identity/structure lock, and source-optics lock.
LIGHT: one L, one E, physical shadow/background response, T, and optional A.
RENDER: one G plus one D, only when either differs from source.
SKIN: one scale-aware S behavior plus one source-consistent P finish.
QUALITY: one Q repair, only when Q differs from source.
AVOID: only the two or three failures most likely for this source.
```

- Use four lines when `G0 + D0 + Q0`; add RENDER only for a requested grade/device and QUALITY only for an observed repair target. Target `55–110` English words, with an absolute ceiling of `135` for combined edits.
- State identity once. Treat the source itself as the identity card; do not invent a new age, personality, beauty description, lens, or aperture.
- Use positive, observable photographic behavior before negative constraints. Omit recipe codes, audits, confidence, backend notes, and reasoning.
- Compile bare words such as `cinematic`, `editorial`, `HDR`, `film`, `DSLR`, `medium format`, or `smartphone` into visible exposure, palette, tonal, microcontrast, sharpening, and dynamic-range behavior.
- Default to `E0 + G0 + D0 + Q0`. Never mix multiple device profiles, color looks, or quality recipes. Never claim exact manufacturer color science or pixel identity from prompt wording.
- Keep default prompts free of realism-by-dirt terms: freckles, blemishes, blackheads, rough skin, color irregularity, film grain, gritty texture, under-eye lines, and high contrast. Preserve source-specific marks without naming or amplifying them.
- Use `deep`, `near-black`, or `hard contrast` only when the user explicitly selects backlight, hard light, low-key, neon, or silhouette behavior.
- On retry, replace the failed line instead of appending more instructions.

Every handoff must use this structure with no placeholders.

## Preserve the Identity Signature

Keep six source-defined groups stable: face outline/proportions; feature spacing, shape, and size; hairline, parting, and hair mass; source-identifying skin anchors; makeup/accessories; and apparent age/expression. Do not describe these groups in detail to the image model unless a real failure requires a shorter identity retry. Detailed identity checks belong in validation, not the generation prompt.

Under A6, internal features are intentionally hidden. Judge identity from hair/head/body outline, head-to-body ratio, pose, position, and framing.

## Apply Optical Skin Realism

- Keep overall skin color clean and continuous across face, ear, neck, and visible upper chest. Local transitions should be gentle and source-consistent, never patch-like.
- Use a diffuse base with small bounded specular highlights only on planes facing the selected key. Avoid a whole-face gloss layer and avoid removing all highlights.
- Vary surface detail by region: cheek pores softer, nose pores slightly clearer, lip texture separate, eye-area structure undisturbed. Do not tile one pore pattern across the face.
- Match detail to the source focus plane, depth of field, face size, and illumination. Never sharpen the whole face, every hair, clothing, and background equally.
- Preserve source-existing marks as identity anchors, but do not list or generate new imperfections by default.
- Preserve source skin tone; do not use `fair`, `whiter`, or beauty-grade language unless the user explicitly requests a complexion change.

## Apply Physical Light Without Unwanted Darkness

- Default preservation to E0 and natural relighting to E1, with clean midtones and readable but directional shadow separation.
- Add fill only when the selected exposure requires information to remain readable. Do not flatten intended backlight, hard light, low-key, or silhouette.
- Derive shadow edge from apparent source size and distance. Carry direction, falloff, cast shadows, and reflected color across subject, clothing, nearby surfaces, and background.
- Keep window/tree shadows continuous across curvature and adjacent surfaces; keep bokeh only in optically defocused regions; require a visible or strongly inferred source for rays; give neon a clear primary and secondary source.
- Treat prism color as refraction of one existing source, not painted rainbow patches. Give blind/lattice shadows one projection geometry, candlelight rapid near-field falloff, rim light a continuous rear-quarter outline, and rain-wet highlights the same direction as the key.
- Under A6, render the complete subject interior as one clean black mass. Permit only a narrow source-consistent rim that does not enter the silhouette.

## Apply Tone and Capture Style Precisely

- Separate scene light from image rendering. L/T define the physical source; E places highlights, midtones, shadows, and black point; G defines palette and curve; D defines capture response.
- For exposure changes, name all four tonal zones. Do not request simultaneous global highlight recovery and shadow lifting; preserve directional contrast.
- For color/style changes, state skin-neutral placement, neutral-object behavior, saturation relationship, contrast curve, and highlight roll-off. A tint alone is not a style.
- Translate device names into visible behavior. Full-frame, medium-format, 35mm negative, CCD compact, point-and-shoot flash, smartphone computational, instant film, and disposable-camera profiles live in the device reference.
- Device emulation preserves camera position, perspective, crop, focal plane, and depth-of-field strength unless `optical-restyle` is explicit. If optics change, state the visible consequence rather than lens numbers alone.
- Grain, vignetting, borders, color casts, exposure defects, and date stamps are opt-in analog artifacts. They are frame-level effects and never create skin realism.

## Repair AI Generation Quality Without Flattening Detail

- Read [the V2.3 quality reference](references/quality-repair-and-iteration.md) for dirty synthetic patterns, gradient banding, edit seams, color drift, or cumulative multi-round degradation.
- Distinguish synthetic repeating residue from real pores, hair, fibers, weave, film grain, edges, and focus falloff. Remove only the observed defect; never apply generic denoising or whole-frame smoothing.
- Treat banding as broken luminance/chroma continuity, not as a shadow to lift. Preserve real object edges, cast-shadow boundaries, local contrast, and black point.
- For local seams, separate the target, its legitimate interaction pixels, and protected surroundings. Match boundary light, color, sharpness, depth, noise, haze, and reflections without blurring the whole region.
- For color drift, use the first root base—not the previous round—as the color reference. Lock only unchanged or unauthorized regions; never undo an intentional relight, grade, monochrome conversion, or recolor.
- For repeated edits, use the current accepted edit as Image 1 and the root base as Image 2. Repair after every round before continuing, and do not attach the full history.
- Prompt constraints reduce regeneration but cannot hard-lock pixels. Never use local Pillow/NumPy/OpenCV filters to produce the final image or claim deterministic preservation.

## Preserve the Frame

Retain orientation, aspect ratio, composition, focal plane, depth of field, and subject-to-frame scale. A backend may downscale uniformly; exact pixel dimensions are not required. If a local result exists, run the read-only check:

```bash
python3 "$XXG_SKILL_DIR/scripts/check_aspect_ratio.py" SOURCE_IMAGE EDITED_IMAGE
```

Accept relative aspect-ratio drift of `≤5%`. Never resize, crop, pad, or extend locally to force a pass.

## Validate the Result

After generation, read [the V2.3 identity and detail audit](references/identity-and-detail-audit.md). At normal size first, verify:

1. the requested lighting/tone/capture change—or exact preservation of every unauthorized axis—is immediately clear;
2. identity signature, pose, source optics, framing, and subject scale remain stable;
3. skin reads clean before microdetail becomes visible, with bounded highlights and no uniform gloss or texture overlay;
4. detail density follows facial region, focus, scale, and illumination rather than appearing equally sharp everywhere;
5. subject and environment share one physical light system; exposure, grade, and capture response are coherent; A6 remains a complete black interior.
6. the selected Q repair is visible without erased natural detail, unintended color rollback, new seams, or later-round quality loss.

If a result is nearly unchanged, strengthen one observable target. If skin becomes artificial, replace SKIN with `clean continuous source skin tone; bounded source-shaped highlights; faint region-specific microdetail only where focus and light resolve it`. Never present a failed image as final.

## Load References Only When Needed

- For every prompt: `references/prompt-recipes.md` and `references/lighting-skin-color-temperature-recipes.md`
- For tone, exposure, grading, style, device, film, camera, phone, CCD, or optics requests: `references/tone-exposure-style-device-recipes.md`
- For dirty texture, banding, posterization, seams, color drift, or repeated-edit degradation: `references/quality-repair-and-iteration.md`
- For tool routing, backend classification, or failure handling: `references/backend-and-clean-realism.md`
- After generation: `references/identity-and-detail-audit.md`
- Only for a verified strict local-edit backend: `references/edit-plan-and-protection.md`