comfy-imagedescribe

Use this skill for local image description through the comfy-imagedescribe CLI. The CLI uses the Qwen3-VL 2B Instruct HuggingFace model directory under /mnt/models/comfyui/LLM/Qwen-VL/Qwen3-VL-2B-Instruct. The model is a full vision-language model loaded with transformers (Qwen3VLForConditionalGeneration), not a diffusion text encoder, so the image is actually grounded in the generated description. It is not distributed by comfy-models download; obtain the model directory from Hugging Face (Qwen/Qwen3-VL-2B-Instruct) and place it under LLM/Qwen-VL/ if missing.

The CLI is quiet by default and prints only final JSON. Use --verbose only when debugging ComfyUI runtime output, warnings, or progress bars.

If comfy-imagedescribe or comfy-models is not available, use comfy-tools-setup first. In this repository, prefer uv run comfy-imagedescribe; outside the repo, let comfy-tools-setup install the CLIs with uv tool.