LLM Studio Structured Describe (Video)

The image version of this node gives you clean structured JSON from one picture. LLM Studio Structured Describe (Video) is the same promise for video: it samples frames from a clip, runs them through a local vision model with JSON-schema constrained output, and hands you guaranteed-valid JSON - plus the individual fields pulled out and a count of how many frames it actually processed.

Inputs split into "where's the video" and "how should the model behave." video_path is a path to a video file, and sample_rate + max_duration control how much of it gets analyzed - the same framing as the other video nodes: frames per second to extract (default 2.0) and a maximum duration in seconds (default 5), so a long clip doesn't balloon into a slow, expensive sampling run. Then base_url and model_name (default qwen3-vl-8b-thinking-mlx) point at your LM Studio server, schema_preset picks the JSON schema (video_description, simple_description, or character_analysis), and prompt is the instruction text. temperature defaults low at 0.2, and verbose toggles logging.

Outputs: json_output (the full valid JSON as a string), field_1 through field_5 (individual fields sliced out per the preset - wire one straight into a prompt builder), and frames_processed (an INT telling you how many frames were actually sent, which is your sanity check that sampling behaved).

Why reach for this over the free-form Video Describe? If the destination is any kind of automation - a logging pipeline, a prompt-builder that expects named fields, a batch job - structured output saves you a parsing step and the JSON-schema enforcement means the shape never breaks between runs. The free-form node is friendlier to read; this one is friendlier to build on. The smaller default model (8B vs 30B) also makes it the budget option for iterating.

Installing

Ships with 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

• LM Studio needs JSON-schema support for the loaded model, same as the image variant. Schema errors mean an outdated server or a model without structured-output support.
• sample_rate × max_duration is your cost dial. 2 fps × 5 seconds = 10 frames per video. More frames means slower runs and more tokens; start low and raise it only if the captions are missing detail.
• frames_processed is your friend. If it's 0 or suspiciously low, the video path is wrong or the sampler didn't find frames - check the path before blaming the model.
• Fields are sliced per preset, so change the preset and field_1..5 mean different things.

This is the node for when you want video captions that slot into code or other nodes without ceremony. Local, structured, and no cloud key anywhere in the chain.