Load Model (VideoChat)

VCW_ModelLoader is where the video-chat half of lebrosoft/ComfyUI-VideoChatWrapper actually gets heavy. This node loads the model that understands videos - a ~7B vision-language model called VideoChat-R1, the 7B variant from OpenGVLab (the same lab behind the InternVL video-understanding line). It's a fine-tune of Qwen2.5-VL-7B-Instruct, Apache-2.0 licensed, built with reinforcement learning so it reasons about video instead of just describing frames. And it's genuinely obscure: on HuggingFace the model repo has a few hundred downloads. You are an early adopter whether you planned to be or not.

How it works

There's exactly one choice in the dropdown - VideoChat-R1_7B - and the node resolves it to OpenGVLab/VideoChat-R1_7B on HuggingFace. On the first run it downloads the whole checkpoint to ComfyUI/models/videochat/VideoChat-R1_7B/ (a non-standard folder, so don't go looking for it in the usual model directories) and then loads it with HuggingFace Transformers: Qwen2_5_VLForConditionalGeneration, torch_dtype="auto", device_map="auto". It caches the loaded model on the node instance, so re-running your workflow doesn't reload 15 GB from disk every time - but a ComfyUI restart starts from scratch.

The two outputs are important to understand:

• model - typed as MODEL, but this is not a diffusion checkpoint. It's a Transformers vision-language model. Do not wire it into a KSampler; that's the most common way to confuse yourself with this pack.
• processor - the paired AutoProcessor that tokenizes and resizes inputs for the model.

Both feed into VCW_VideoSummary.

What it costs you

The 7B model loads in bf16, which puts it around 14–16 GB in VRAM depending on context. You want a 16 GB+ card to run it comfortably. device_map="auto" means it will run on less by offloading layers to system RAM - technically possible, but you'll be waiting a long time per summary. Plan for the full model or plan to wait.

Troubleshooting

The failure modes here are the standard local-transformers ones, and the pack's README won't help you with any of them:

• First run needs internet and disk. It's a silent ~15 GB download that starts the moment you queue. If it fails mid-download, the partial checkpoint in models/videochat/ can leave the load stuck - delete that folder and retry.
• transformers version. Qwen2.5-VL support needs a reasonably recent Transformers. The pack's requirements.txt lists only qwen_vl_utils (and a commented-out transformers), so if your ComfyUI environment has an old Transformers, the import or load will fail. pip install -U transformers qwen_vl_utils is the usual cure.
• VRAM OOM - see above.

Installing it

With the pack: ComfyUI Manager → search ComfyUI-VideoChatWrapper → install, or:

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

Restart ComfyUI. No manual model download - the node handles that - but budget disk space and patience for the first queue. One light aside: for a pack this small and quiet, a 7B auto-downloading model is the definition of "the hidden cost is in the first run."