still-qa
Installation
SKILL.md
still-qa
python scripts/judge.py --endpoint $VLM_ENDPOINT --model $VLM_MODEL --image out/s03.png --prompt "<shot prompt>" --bible bible.md [--ref cast/mara_front.png]
This prints JSON with scores, mean, verdict (pass/retry), issues and fix.
The loop
- Generate the still, then judge it.
- If the verdict is
pass(every score 3 or higher and a mean of 3.5 or higher), accept it. - If the verdict is
retry, apply the singlefix: a prompt edit, a new seed, or animage-editpass. Regenerate and judge again. Allow at most 2 retries, then show the best take to the user with its issues. - Log every take's scores in the plan so
continuity-checkcan use them.
Judge setup
- Serve an open-weights VLM (Qwen2.5-VL-7B-Instruct, Apache-2.0, unverified for your build) with vLLM, llama.cpp server, Ollama or LM Studio. Any OpenAI-compatible
/v1/chat/completionsendpoint works. A 7B VLM needs about 16 GB in bf16, or about 6-8 GB quantized (estimate, unverified). - In Comfyrizz-style setups, the orchestrator's own judge (for example Pi) can replace
judge.pyas long as it keeps the same rubric and JSON shape. - The judge must see the prompt and the bible descriptors, not only the image.
- For video, judge 3 frames: the first, the middle and the last.