Video Summary (VideoChat)

VCW_VideoSummary is the payoff node of lebrosoft/ComfyUI-VideoChatWrapper: the one that actually watches a video and writes a summary of it. It takes the model and processor loaded by VCW_ModelLoader plus the file path from VCW_LoadVideo, runs the whole thing through a ~7B vision-language model, and hands you back a chunk of text. If you've wanted ComfyUI to look at footage and tell you what happened, this is the pack's entire reason to exist.

How it works

The node builds a chat message the way Qwen2.5-VL expects: a video content block pointing at your file, plus a fixed user prompt - "Describe this video in detail." - then it samples frames, applies the model's chat template, generates text, and returns it. A few hardcoded choices shape the result and are worth knowing:

• fps = 1.0 - the model sees roughly one frame per second of footage. Good for an overall summary, hopeless for frame-by-frame detail.
• max_pixels = 360 × 420 - frames are scaled way down. Fine for "what is this video about," bad for reading text on screen or counting small details.
• max_new_tokens = 512 - the summary is capped at 512 generated tokens. Long videos get a truncated description, not an exhaustive one.

The prompt is also hardcoded. There is no prompt input on this node, so you can't ask "what color is the car in scene two?" - you get a detailed description, take it or leave it. That's the honest limit of this node: it's a summarizer, not a chat interface.

Inputs and output

• model - from VCW_ModelLoader.
• processor - from VCW_ModelLoader.
• video_path - the STRING path from VCW_LoadVideo (or any node that produces a video path).
• text - the output: a STRING containing the generated summary. Wire it into a text display node - the pack's demo workflow uses pysssss's ShowText - or into anything that consumes a string.

Gotchas

• CUDA is assumed. The node sends inputs straight to "cuda". No CUDA GPU, no summary.
• The first run is a marathon. It downloads the model (see VCW_ModelLoader), then processes your whole video at 1 fps. A long clip means a long wait and a lot of VRAM.
• No prompt control. Again: you get a description, not a Q&A. If you want to actually chat about the video - ask follow-ups, query specific moments - this node isn't the tool, and the pack doesn't provide one.

Installing it

With the rest of the pack: ComfyUI Manager → ComfyUI-VideoChatWrapper, or:

cd ComfyUI/custom_nodes
git clone https://github.com/lebrosoft/ComfyUI-VideoChatWrapper

Restart, then queue once to trigger the ~15 GB model download to ComfyUI/models/videochat/. The real prerequisites are the ones the README won't emphasize: a 16 GB+ GPU, a recent transformers (pip install -U transformers qwen_vl_utils if the load errors), and patience for that first run.

Bottom line: this is a niche, quietly-maintained pack, and the summary quality lands in "competent local VLM" territory - real, occasionally surprising, not frontier. If you need one-shot video descriptions without leaving ComfyUI, it works. If you need a video chatbot, keep looking.