LLM Studio Picture Describe

LLM Studio Picture Describe is the workhorse of the pack's local captioning family: you feed it one image tensor and it returns one caption paragraph from a vision model running in LM Studio. No API key, no cloud round-trip - just you, LM Studio, and a model on your own GPU.

The inputs follow the LLM Studio pattern. base_url and model_name point at your LM Studio server, image takes the actual IMAGE tensor (from a Load Image, a Frame Extractor output, wherever), and caption_prompt is the instruction sent to the model. The default prompt is worth reading because it reveals what this node is tuned for: a detailed hair-and-facial-features caption - texture, style, length, facial structure, eyes, brows, nose, lips, skin tone, markings, and how light falls on the face. That's character-consistency territory. If you want reusable descriptions of a person so a face survives across generations, this default is aimed right at that job. If you want a general caption instead, replace it with your own text.

Two controls matter beyond the obvious. api_endpoint offers chat/completions and completions - and the tooltip is the truth: chat/completions may include thinking/reasoning tags in the output, while completions is cleaner. For caption text you'll actually paste into a prompt, completions is usually the one. And temperature defaults to a low 0.2, which is right for description - you want consistent, deterministic captions, not creative reinterpretations. verbose toggles console logging when something's misbehaving.

Mechanically it's plain: the image gets encoded and posted to LM Studio's OpenAI-compatible endpoint, and the returned text comes out as a single STRING output, caption. From there it can feed a prompt builder, a logging node, or a Multi-Caption Combiner when you're assembling per-frame descriptions of a clip.

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 .

And of course: LM Studio running with a vision model loaded on port 1234.

Gotchas

• Vision models only. A text-only model returns gibberish or an error. You need a VLM like a Qwen-VL variant.
• The default endpoint is completions for a reason. If captions come back littered with 思考...-style thinking tags, you're on chat/completions with a reasoning model - switch endpoints or the model.
• The default caption_prompt is very face-focused. Great for character work, wrong for "describe the scene" use. Don't treat it as a general default; own your prompt.
• A low temperature is your friend here. Description wants consistency, not creativity.

It's the node you reach for when you want a specific, reproducible read on one image without paying a cloud API. Pair it with Frame Extractor and Multi-Caption Combiner and you've got a fully local video-captioning chain that never touches the internet.