LLM Studio Video Describe

Most video models are bad at "watch this whole thing and tell me what happened." They sample a handful of frames and stitch a guess. LLM Studio Video Describe doesn't hide that - it makes it the workflow. It samples frames from a clip at a rate you set, captions each one with a local vision model, and hands you both the per-frame captions and a combined narrative. You keep the detail or you take the summary; both are yours.

Inputs: video_path for the clip, sample_rate (frames per second, default 2.0) and max_duration (seconds, default 5) to control sampling, base_url and model_name (default qwen/qwen3-vl-30b) for LM Studio, api_endpoint (completions by default for clean output, or chat/completions if you don't mind thinking tags), temperature (default 0.2), and verbose. Then there's caption_prompt - the instruction sent for each frame - and this is where you need to be an adult about defaults: the one shipped in the pack is aggressively tuned for explicit adult-video captioning. Swap it for something neutral before you caption anything you care about; this node doesn't care what you ask it, it just asks faithfully.

Outputs are the useful part: combined_caption (the merged description of the whole clip), frame_captions (each individual frame's caption), and frames_processed (an INT - how many frames actually went through, your sanity check on sampling). The combined_caption is what feeds a prompt builder for a re-render; frame_captions is what you'd feed a Multi-Caption Combiner if you want a differently-styled merge.

Mechanically it's the classic sample-and-caption loop: extract frames via OpenCV, base64 them into an OpenAI-compatible call to LM Studio, collect the text. The output is only as good as your prompt and your model - a 30B Qwen-VL handles detail well but runs slow; an 8B is faster and fuzzier.

Installing

Part of Swiss Army Knife:

• ComfyUI Manager: search "Swiss Army Knife", install, restart.
• Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/sammykumar/ComfyUI-SwissArmyKnife.git
cd ComfyUI-SwissArmyKnife
pip install -e .

Gotchas

• The default caption_prompt is NSFW-tuned. Change it first. This is a general-purpose captioner hiding behind an adult-workflow default.
• sample_rate and max_duration govern your time budget. 2 fps over 5 seconds is 10 frames; raise either and the run gets proportionally slower.
• Frames are captioned independently. Each frame is a separate call, so consistency across frames depends on the model - don't expect it to hold a narrative thread without help.
• If frames_processed comes back 0, the video path is wrong or unreadable. Fix the path, not the node.

It's the free-form sibling of the structured video node - read the two side by side and pick by whether your downstream wants JSON fields or a paragraph.