RedNode Describe To Boxes

The Krea 2 / prompt-frame approach this pack pushes splits a prompt into Subject and Surroundings because word order sets the framing - open with the person and you get a close shot, open with the room and the camera pulls back. The laborious part is writing those boxes from a reference image by hand. RedNode Describe To Boxes does that for you: it sends a picture to a local Ollama vision model, gets back five labelled sections, and hands Subject, Surroundings, and Light-and-colour out on separate sockets, ready to wire into a RedNode Prompt Frame.

Mechanically it's clean. You feed an image, pick the engine (Ollama is the only verified one - the tooltip notes the ComfyUI QwenVL nodes currently can't execute), pick the model from what Ollama has (tested on qwen3-vl 8b instruct at Q4), and hit run. The node asks for five sections - SUBJECT, SURROUNDINGS, LIGHT, COLOUR, MOOD - each on its own line, and parses the reply into the labeled outputs. There's a raw output carrying the full reply in case the split comes back malformed, and a notice output for when something went wrong (no model found, Ollama not running).

Three details are worth knowing before you use it. First, the instruction text is the whole game: the default wording is the one that scored 100% on format across 30 captions, and the parser keys on the exact section names - change them and the split breaks. Leave it alone unless you know what you're doing. Second, max_tokens is the reply cap, and too low truncates MOOD - the last section, so it dies first. Third, seed makes the description reproducible: same seed, same caption.

The best design choice here is what happens after the reply: the vision model is released from VRAM the moment the response lands (keep_alive zeroed), so it doesn't sit on top of your checkpoint for the rest of the queue. That's the classic low-VRAM trap with LLM nodes - running two models at once because nobody unloaded the first - and this node refuses to fall into it.

One thing the wider ecosystem should teach you before you get too trusting: any LLM/VLM node is arbitrary Python reaching the network, and that exact category already shipped malware once (the ComfyUI_LLMVISION incident). This node is one of the good shapes - a local Ollama call, no bundled model downloads, no hidden API keys - but "local and open" is precisely why it's the right instinct.

Installation is the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/RedNodeAI/ComfyUI-RedNodeStudio.git

or search RedNode Studio in ComfyUI Manager. You'll also need Ollama running locally with a vision model pulled (ollama pull qwen3-vl:8b), which is the one external dependency. No pip requirements beyond that. Works with any model - the Ollama side doesn't care what you generate with.