ComfyUI Node

Video Chat 3

Analyze sampled visual frames with VideoChat3-4B. Model path, precision, download, and attention settings are configured directly on this node. Audio is not transcribed. GPU memory is released after execution.

By endman100·Created 2 months ago·Updated 2 months ago· 1
Video Chat 3
  • video
  • Final answer
  • Analysis JSON
  • Checkpoint path
◄modelhf://MCG-NJU/VideoChat3-4B►
◄promptCreate a timestamped summary and identify important visible events.►
◄chunk_seconds120►
◄nframes8►
◄max_pixels100352►
◄segment_tokens192►
◄summary_tokens512►
◄chapter_size12►
◄model_path►
◄precisionauto►
◄attention_backendauto►
◄allow_downloadtrue►
◄revision37fa901ec5913f84bc31108ebc1e60ad1903634c►
◄resumetrue►
◄output_prefixvideochat3/analysis►
Categoryvideo/understanding

Inputs (16)

NameTypeDefaultDescription
videoVIDEOConnect ComfyUI Load Video or another VIDEO-producing node.
modelCOMBOhf://MCG-NJU/VideoChat3-4BHugging Face cache or a model discovered in ComfyUI's registered videochat3 paths.
promptSTRINGCreate a timestamped summary and identify important visible events.Question or analysis goal. VideoChat3 analyzes visuals only, not audio.
chunk_secondsFLOAT12010–3600—
nframesINT82–64—
max_pixelsINT100352100352–602112—
segment_tokensINT19232–1024—
summary_tokensINT51264–2048—
chapter_sizeINT122–64—
model_pathSTRINGOptional absolute path or path relative to a registered videochat3 model root. Overrides model.
precisionCOMBOauto4 options: auto, bf16, fp16, fp32
attention_backendCOMBOauto4 options: auto, sdpa, eager, flash_attention_2
allow_downloadBOOLEANtrueAllow the pinned Hugging Face model revision to download during execution.
revisionSTRING37fa901ec5913f84bc31108ebc1e60ad1903634cPinned model-code revision. Changing it executes different remote Python code; do not trust revisions from unknown workflows.
resumeBOOLEANtrue—
output_prefixSTRINGvideochat3/analysis—

Outputs (3)

NameTypeDescription
Final answerSTRING—
Analysis JSONSTRING—
Checkpoint pathSTRING—