SKILL

When to use

Setting up an open-weight model in ComfyUI on Windows when VRAM is below the official minimum or the model needs a newer core. Confirm facts, protect the working env, use a community low-VRAM recipe, verify weights, then prove it with a boot self-test plus (image models) a real smoke image.

Steps

1. Confirm via web, not memory: license, official HF repo (beware fakes), real checkpoint, min ComfyUI version, and the EXACT text encoder + VAE (generation N often differs from N-1).
2. Probe host: nvidia-smi --query-gpu=name,memory.total --format=csv, RAM, disk, arch (NVFP4 needs RTX50). Probe existing ComfyUI version, custom_nodes, yaml.
3. Never upgrade a working env in place: robocopy the portable to a new dir, upgrade the copy with update/update.py (needs .git).
4. After core upgrade re-run pip install -r requirements.txt -i https://pypi.org/simple to fix stale deps.
5. Recipe: GGUF/INT8 DiT + BlockSwap + quantized text encoder + community sweet-spot resolution/frames/steps. Watch quant quirks (dim not divisible by 256 -> Q4_0/Q5_1).
6. git clone --depth 1 nodes into custom_nodes; install reqs with embedded python.exe.
7. List exact filenames/sizes with HfApi().model_info(repo, files_metadata=True) before pulling GB.
8. Weights on a roomy drive via a dedicated extra_model_paths.yaml. CJK paths: ASCII -File script with [char]0x...., results to a file. Resumable download via mirror. 401 or ~150-byte ‘Access restricted’ = gated; find a re-upload.
9. Boot test: --lowvram --listen 127.0.0.1 --port <p> --disable-auto-launch in background, logs, wait ~45s, grep for IMPORT FAILED/Traceback, low-VRAM state, paths. If a watchdog says stalled, check the port/API before relaunching.
10. /object_info must list your filenames; then a /prompt smoke image, poll /history. Leave OOM-prone video smoke tests to a human pass.
11. Some nodes ignore the yaml (grep folder_paths/models_dir); restart after adding weights.
12. Write a launcher .bat + deploy manual.

Gotchas

• mat1 and mat2 shapes cannot be multiplied at txt_in = WRONG text encoder file (e.g. Qwen3 vs Qwen3-VL). Copy CLIP type, sampler and encoder name from the official template JSON in comfyui_workflow_templates_json/templates/; don’t guess.
• aria2 -i input files mangle CJK paths (errorCode=18): pass -d/-o as argv, one file per call, one process only.
• -like '*name*' process checks match your own shell.
• hf_hub_download may throw LocalEntryNotFoundError while curl works; use a resumable downloader. Verify size and no .aria2 file.