LLM Studio Structured Describe (Image)

The annoying thing about asking an LLM for JSON is that you often get a JSON-ish paragraph with commentary glued to it, and then you're writing parsers and praying. LLM Studio Structured Describe (Image) removes that entire failure class: it uses LM Studio's JSON-schema structured output mode, so the model is constrained to return valid JSON matching a schema you pick. No code fences, no "here is the output:" preamble, no manual parsing.

You feed it an image tensor, a base_url and model_name for your LM Studio server, a prompt, temperature (default 0.2), and schema_preset. The preset is the heart of it - three choices, each a different JSON schema:

• video_description (default) - the full structured description, matching the pack's five-paragraph style
• simple_description - a lighter caption with fewer fields
• character_analysis - aimed at breaking down a character, which pairs with character-consistency workflows

On the way out it returns json_output (the complete valid JSON as a string) plus field_1 through field_5 - the individual fields pulled out as separate STRING outputs. That's the part people underrate: you don't have to parse the JSON to get at one field. If you only care about the subject, wire field_1 straight into your prompt builder and ignore the blob.

Mechanically this is where LM Studio's structured output support does the real work - the node sends response_format as a JSON schema and the server enforces it, which is a different and much more reliable path than hoping the model behaves. The default model here, qwen3-vl-8b-thinking-mlx, is notably smaller than the 30B used by the free-form Picture Describe - a deliberate choice, because structured output doesn't need the biggest model and 8B is fast enough to iterate on.

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 must support JSON-schema structured output for the model you loaded. Newer LM Studio versions do; if you get schema errors, update LM Studio or the model.
• schema_preset defines which fields come out, and field_1..5 are sliced according to that preset. Change the preset and the field meaning changes - the node doesn't re-label them for you.
• Structured output guarantees shape, not quality. Valid JSON can still be mediocre content. If the captions are wrong, that's a prompt or model problem, not a schema one.

If you're building any automated pipeline off local captions - feeding a prompt builder, logging structured descriptions, driving character consistency - this is the node that keeps the data clean. Free-form captions are friendlier to read; structured output is friendlier to build on. Pick accordingly.