| name | image-gen |
| description | Use this skill to create, generate, draw, or edit images/photos into saved PNG files with OMLX image models. Covers text-to-image and image-to-image edits. Do not use for text-only image explanations, screenshot summaries, or non-generative image processing scripts. |
| allowed-tools | Bash(uv run ./scripts/generate.py:*) Bash(uv run ./scripts/edit.py:*) |
OMLX Image Generation and Editing
Requirements
$OMLX_BASE_URL must be set.uv must be available to run the bundled Python helpers.- Editing requires a source
.png, .jpg, .jpeg, or .webp file.
- If the loaded task is non-generative image analysis or simple file processing, stop using this skill and handle it with ordinary tooling.
Workflow
- Keep user inputs and outputs in the user's workspace. Pass absolute workspace paths for
--input, --inputs, --mask, and --output.
- Discover an available model from
$OMLX_BASE_URL/v1/models/status unless the user specified one. The response puts models under .models (not .data). Prefer models with engine_type or model_type of image and a loaded status. A working discovery query:
curl -s "$OMLX_BASE_URL/v1/models/status" | jq '.models[] | select(.engine_type == "image" or .model_type == "image") | {id, loaded, capabilities}'
Prefer loaded models whose capabilities or tasks include generation or edit.
3. Use uv run ./scripts/generate.py or uv run ./scripts/edit.py. Do not hard-code install locations, copy scripts elsewhere, or run non-bundled wrappers. The scripts refuse outputs inside the skill directory.
4. For structural edits that transform a scene while preserving the subject, prefer FLUX.2/Klein edit models over ERNIE when both are available. Prompt explicitly for preserved subject details (pose, clothing, crop, camera angle, privacy blur) and the replacement scene. If the first result changes too much or too little, retry with --image_strength values such as 0.2, 0.4, or 0.6, and use --steps/--guidance for quality and prompt-adherence sweeps.
5. Report the saved file path and model used.
Commands
Generate:
uv run ./scripts/generate.py \
--prompt "a watercolor fox in a moonlit forest" \
--model "<image-model>" \
--size "1024x1024" \
--output "/absolute/path/to/user/workspace/result.png"
Edit:
uv run ./scripts/edit.py \
--input "/absolute/path/to/user/workspace/source.png" \
--prompt "add sunglasses" \
--model "<image-edit-model>" \
--output "/absolute/path/to/user/workspace/edited.png"
Read references/generation.md or references/editing.md only when you need optional parameters such as --n, --quality, --style, --mask, --steps, --guidance, --image_strength, multi-input edits, or URL responses.
