AI视频理解

The image version of this node reverse-prompts a single frame. This one reverse-prompts a whole clip. Feed it a path to a local video file and Alibaba's Qwen-VL model watches the thing and returns a text description of what's happening - camera movement, subject changes, scene, the works. For anyone building image-to-video or trying to match the style of an existing shot, that's a genuinely useful trick, and it costs you zero local VRAM.

The mechanism is blunt but effective: the node reads the whole video file, base64-encodes it, and sends it to DashScope's OpenAI-compatible endpoint as a data:video/mp4 URI. The text box is your instruction (the default asks the model to reverse-prompt the video), mode sets how verbose the answer is (默认, 简短, 详细), out_language picks English or Chinese, and seed makes repeated runs reproducible. You get one output: text, ready to wire into a prompt or another node.

The inputs you actually touch: video_url (a local file path to an mp4 - not a web URL, not a video batch, just a path on disk), model (your choice is qwen-vl-max, qwen-vl-max-latest, or qwen-vl-plus - there's no qvq or 7B option here, unlike the image node), and the key + base_url pair shared by every node in this pack.

That base_url is pre-filled as https://dashscope.aliyuncs.com/compatible-mode/v1 and the api_key field will even pre-fill itself if you've set the DASHSCOPE_API_KEY environment variable.

Install: ComfyUI Manager → "ComfyUI-My-Nodes", or git clone https://github.com/Tagbliton/ComfyUI-My-Nodes into custom_nodes and restart. Dependencies come from the pack's requirements.txt (dashscope, openai, requests, soundfile, numpy, and the famously mandatory urllib3==1.26.20 pin).

Where people get burned - and this is the big one: the README's error list spells it out. Exceeded limit on max bytes per data-uri item: 10485760 means your video, base64-encoded, blew past the API's 10 MB ceiling. The cause is almost always a high-bitrate encode, not a "too long" video per se - a short 4K/60 clip can exceed it while a long, well-compressed 720p one passes. Re-encode the file with a lower bitrate (handbrake or ffmpeg with -crf 28-ish settings) before feeding it in. Everything else is the usual DashScope drill: Connection error means network or proxy trouble, and a key that fails with Invalid Api-Key provided usually means you don't have a real-name-verified mainland China account, which is the actual requirement Alibaba enforces.