MiniMax H3 Enhance Prompt

The blank-page problem is real with MiniMax H3. The Director wants a 350–500 word prompt in MiniMax's guide notation, and what you actually have is "a baker opening a bakery at dawn, a Samoyed walks in." MiniMax H3 Enhance Prompt is the shortcut: you give it reference images plus a one-line idea, it asks a local vision model to write MiniMax-shaped prose, and it hands the same images out on ref_images - so what the model described is literally what H3 conditions on. No cloud, no API key for the default path, nothing sent anywhere you don't point it.

How it works

It's a thin client for an OpenAI-compatible chat endpoint. The reference images get downscaled to max_image_size (768 on the long edge by default), JPEG-encoded, and sent alongside your idea. The model's answer is trimmed to max_words - the guide's 350–500 range - and cleaned of preamble so you don't get "Here is your enhanced prompt:" as literal prompt text. One non-negotiable: it has to be a vision model. A text-only model will ignore your images without saying so, and you'll get a generic prompt that has no idea what your reference looks like. qwen2.5vl:7b under Ollama is a reasonable starting point; larger VLMs write noticeably better prose.

The settings that matter

• idea - what you want, in plain words. Everything else is scaffolding around it.
• preset - global writes scene, style, subjects and lighting and leaves the shots to your timeline. storyboard writes the whole shot sequence with timestamps, which you should only use when your timeline segments carry no prompt text - otherwise the two shot numberings collide.
• provider / base_url / model - Ollama (default http://127.0.0.1:11434), LM Studio, or any OpenAI-compatible endpoint. Host and port only; no path.
• seed - ComfyUI caches node outputs, so an unchanged input means the model is never asked again. Bump it to force a fresh answer.
• unload_after - drops the vision model from VRAM when it's done, so it isn't resident while H3 samples. Leave it on; reloading costs seconds, an OOM costs the render.
• on_error - passthrough hands your raw idea on with a warning, so a stopped Ollama doesn't kill the whole run.
• api_key_env - for a hosted endpoint, this takes the name of an environment variable holding the key (MINIMAX_DIRECTOR_VLM_API_KEY, then OPENAI_API_KEY fall back automatically). Never the key itself - widget values are saved inside the workflow, so a key typed here would travel with every copy you share.

What it outputs, and where it goes

Three outputs, all aimed at the Director:

prompt           → Director.global_prompt
ref_images       → Director.ref_images
duration_seconds → Director's `duration` socket (the connection-only one, in seconds)

duration_seconds is just the number you set passed through, so you only type it once and the model's timestamps match the Director's window.

Wiring and gotchas

The node deliberately does not write section labels, <Picture N> numbering, or shot markers. The Director compiles the structured prompt and assigns the reference ordinals itself - a second set from the model would nest structure inside structure and collide with the Director's numbering.

If the VLM and H3 share a GPU, the vision model is evicted after each run (that's unload_after). Ollama has no per-request device selection, so to run it on a second card you set CUDA_VISIBLE_DEVICES on the Ollama service, not here. Expect 15–45 s per run, during which the queue is blocked.

And set expectations honestly: a local enhancer removes the blank-page problem, it doesn't write the movie for you. Small vision models drift on multi-subject scenes (mixing up who's wearing what), so treat the output as a first draft you read before rendering - the same discipline the KB's LLM-in-ComfyUI essay hammers on for every local prompt enhancer.