ImageGen

imagegen

imagegen is a local image-generation skill that leverages an NVIDIA GPU in your system for OpenClaw.

It uses:

• an NVIDIA GPU with CUDA available through PyTorch
• a local Python runtime
• Hugging Face diffusers
• the black-forest-labs/FLUX.2-klein-4B model

The generated image must be written under:

$HOME/.openclaw/media/outbound/

for the Agent to be able to handle the images and embed them elsewhere.

Files

• SKILL.md: skill instructions and usage rules
• run_image_model.sh: shell wrapper used by the skill
• run_image_model.py: Python entrypoint that loads the model and saves the image

Requirements

Before using this skill, make sure the host has:

• Python 3.10+
• bash
• an NVIDIA GPU
• working CUDA support in PyTorch
• internet access on first run to download model weights from Hugging Face

Python packages required by this skill:

• torch
• diffusers
• transformers
• accelerate
• safetensors
• huggingface_hub
• sentencepiece
• Pillow

Preferred Install: .venv

This is the recommended setup.

It matches the current wrapper, which expects a Python runtime in:

$HOME/.venv/bin/python

1. Create the virtual environment

python3 -m venv "$HOME/.venv"

2. Activate it

source "$HOME/.venv/bin/activate"

3. Upgrade packaging tools

python -m pip install --upgrade pip setuptools wheel

4. Install PyTorch with CUDA support

Install the CUDA-enabled PyTorch build appropriate for the machine. Example:

pip install torch --index-url https://download.pytorch.org/whl/cu124

If the host uses a different CUDA build, use the matching PyTorch install command instead.

5. Install the remaining dependencies

pip install diffusers transformers accelerate safetensors huggingface_hub sentencepiece Pillow

6. Verify the environment

"$HOME/.venv/bin/python" -c "import torch; import diffusers; print(torch.cuda.is_available())"

If this prints True, the runtime is ready.

Alternative Install: local Python, no virtual environment

This option installs packages into the local Python environment instead of a dedicated .venv.

Important note

The current run_image_model.sh script is configured to use:

$HOME/.venv/bin/python

If you install dependencies without a virtual environment, you must also update run_image_model.sh so VENV_PYTHON points to your local Python, for example:

VENV_PYTHON="/usr/bin/python3"

1. Upgrade packaging tools

python3 -m pip install --upgrade pip setuptools wheel

2. Install PyTorch with CUDA support

Example:

python3 -m pip install torch --index-url https://download.pytorch.org/whl/cu124

Use the correct CUDA build for the host if it differs.

3. Install the remaining dependencies

python3 -m pip install diffusers transformers accelerate safetensors huggingface_hub sentencepiece Pillow

4. Verify the environment

python3 -c "import torch; import diffusers; print(torch.cuda.is_available())"

Example usage

$HOME/.openclaw/workspace-pixy/skills/imagegen/run_image_model.sh \
  flux2-klein-4b \
  "A sunset over the ocean" \
  $HOME/.openclaw/media/outbound/flux2-sunset.png \
  1024 \
  768 \
  4 \
  42

Notes

• The first generation may take longer because model weights may need to be downloaded.
• The wrapper creates and uses cache directories under $HOME/.cache/.
• The skill currently supports one model only: flux2-klein-4b but it can easily modify to leverage other models. Just ask your Agent.