coco

Type a rough idea - "astronaut, glowing forest, alien planet" - and this node turns it into a fully-formed prompt written the way your actual generator expects: Danbooru tags for Illustrious, natural language for FLUX.2, motion verbs for Wan, the whole [Shot 1] At 00:02.400 ritual for MiniMax H3. It does this with a local LLM (LM Studio or vLLM), no cloud, no API key of your own. It's the "LLM as a node" pattern that went from novelty to routine in this ecosystem, and it fits the 2026 architecture neatly: if your checkpoint's encoder is itself a language model reading an instruction, having a second LLM write that instruction is just translation between two things that speak the same language.

First, a quirk that will save you ten minutes of searching: the node's display name in the Add Node menu is literally "coco" (under the LLM Prompt Studio category). The README calls it "LLM Prompt Studio (LM Studio / vLLM)", the code says "coco" - the code wins. Search "coco" and you'll find it.

How it works

It's a chat-completion client pointed at an OpenAI-compatible server. Set base_url (LM Studio: http://localhost:1234/v1, vLLM: http://localhost:8000/v1), leave api_key as anything for LM Studio or your --api-key for vLLM, and leave model empty - the node auto-detects the chat model loaded at that address (it skips text-encoder and embedding models). The 🔄 button just shows you which one it found.

The interesting part is the cards. The target_model dropdown holds twelve prompt presets - Anima base v1, Illustrious, SDXL, FLUX.2 Klein, Krea 2, Ideogram, LTX-2, Wan 2.2, three MiniMax H3 variants, and a generic one. Pick a target and its English prompt "card" loads into system_prompt, fully editable. Your own rules go in global_directives, which stay put across target switches. user_prompt is your idea. Outputs: prompt (the cleaned text - no thinking, no chat scaffolding) and raw_response (the untouched answer, for debugging).

The reasoning-model handling is the most thought-out bit. Chat LLMs love to emit a <think> block, and that block bleeding into your conditioning is exactly the failure mode the KB warns about. thinking defaults to off and, for reasoning models, states it in every dialect at once - /no_think for Qwen3, enable_thinking=false for GLM, a non-thinking model alias for DeepSeek - then strip_before_tag (default </think>,</mm:think>) cuts any surviving reasoning block out of the output. The model writes like a writer, not like a chat bot.

The inputs a beginner actually sets

• target_model - the generator you're prompting for. Everything else follows from this.
• user_prompt - your idea. The one text box you'll touch every run.
• temperature / max_tokens - 0.7 / 1024 are sane defaults; lower the temperature if output starts repeating.
• image (optional) - connect one and a vision model can caption or describe it into the prompt, labelled <Picture 1>.

There are up to 8 image sockets, plus video (video with video_stride, video_max_frames, video_frame_size, video_fps, video_role) and audio (audio + audio_role), all speaking MiniMax H3's reference vocabulary. image_analysis_size (default 1 MP) downscales before sending - worth turning down once you're feeding several images.

Install & quick start

ComfyUI Manager: search comfyui-llm-prompt-studio. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/grandju51/comfyui-llm-prompt-studio

Restart, refresh. No pip install - the pack is standard-library only, so nothing to break. For LM Studio, load a model and hit Start Server in the Developer tab. For vLLM:

vllm serve Qwen/Qwen3-8B --port 8000

The traps

The one that actually OOMs people is LM Studio's context reservation. context_length × context_slots is memory the model books next to its weights, and LM Studio's defaults (e.g. 60k tokens × 4 slots) can take a 24 GB card out in the middle of a prompt. The symptom is Channel Error in the LM Studio log and {"error":"terminated"} here. A picture plus a few turns of history fits in ~8k, so set context_length around there and context_slots to 1 (ComfyUI sends one request at a time - 4 slots is paying for nothing).

Second trap: min_p and top_k are non-OpenAI extension fields, and some backends answer 400 on the first one they don't know. A vLLM server running speculative decoding always refuses min_p. The node retries once without all extensions and prints what it dropped, and the two enable_* switches let you park tuned values while keeping them out of a request the backend would reject. If you see a request that looks "cancelled," this is usually it.

Leave unload_after off while you iterate - reloading a big model costs seconds per run. Turn it on once your prompt is right and you want the VRAM back for the Flux/SDXL model downstream.