MiniMax H3 GGUF Prompt Enhancer

MiniMax H3 is genuinely good at following what you write - the community reaction when it dropped was "it followed the prompt perfectly." The catch is that "what you write" means its strict sectioned contract, not whatever sentence you typed into a text box. Simple prose gets you audio desync and dialogue in the wrong language. This node is the local, offline answer: it runs a GGUF text model through a private llama-server and returns an H3-compliant prompt plus calibrated width/height - no API key, no cloud, nothing downloaded behind your back.

What this node actually is

Don't be confused by the name: the GGUF here is not H3. H3 is the 33B audiovisual diffusion model that renders video. This node is a separate, much smaller text model - an instruct-tuned chat LLM in GGUF form, like Qwen or Llama - whose whole job is rewriting your rough idea into H3's documented structure (the three-block T2VA contract, or the six-block Ref2VA anatomy when you're feeding references). Think of it as the LLM-as-a-tool pattern from llm-in-comfyui.md: a tool you bolted onto the workflow, built around one specific model's prompt spec.

What makes it different from most local-LLM enhancers is the backend. Instead of loading llama-cpp-python inside ComfyUI's process, it spawns a standalone llama-server binary bound to 127.0.0.1 on a random port with a random API key. That buys crash isolation (a native llama.cpp segfault can't take ComfyUI down), no Python wheel conflicts, and deterministic VRAM reclaim - kill the process and 100% of that VRAM is free before H3 starts sampling.

The inputs that matter

You mostly set four things:

• basic_prompt - your rough idea. Be sloppy here; it's the point.
• mode - auto figures out whether this is t2va, ref2va, i2va, fl2va, l2va, or chained_multishot from your references and wording. Force it when auto guesses wrong.
• gguf_model_path - pick the .gguf from the dropdown. It scans ComfyUI/models/llm_gguf/, your LM Studio cache, and anything in registered_model_dirs.
• llama_server_path - path to an existing llama-server executable. The tooltip is blunt: never downloaded automatically. If it's on your PATH it's found for you.

creative_latitude is the one worth understanding before you crank it. It's a four-step dial for how far the writer may go beyond your words: verbatim_source reformats only; conservative_grounded adds the minimum the mode needs; enhanced_production (the default) resolves unspecified composition, blocking, lighting, micro-performance; invented_production treats your text as a premise and builds the world around it. Quoted dialogue, reference identities, duration, shot count and gore level stay locked at every level.

gpu_layers (auto/all/-1/exact), context_size (default 32768), threads, temperature (0.2), max_tokens (8192) are the llama-server knobs - leave them alone until something's slow or OOMs. repair_attempts (default 2) lets the LLM re-run when validation fails. The big one: keep_server_loaded defaults to false, meaning the server starts, does its job, and dies - freeing VRAM instantly. Set it true only if you're re-running constantly and have headroom, and keep the MiniMaxH3UnloadGGUFServer node handy for when you want it gone.

The outputs

The headliner is enhanced_prompt - wire it into your H3 conditioning. Then width and height, already aligned to 16-pixel steps for your chosen aspect_ratio and target_megapixels, so they feed the empty-latent and sampler nodes directly (1280×720 for 16:9 at auto). validation_report says what was wrong before the repair loop fixed it; enhancement_manifest is the JSON audit trail, and duration_seconds, aspect_ratio, treatment_warnings round out what you'd pass downstream.

Install and first run

Same pack as everything here: ComfyUI Manager (search ComfyUI-MiniMax-H3-Prompt-Enhancer) or:

cd ComfyUI/custom_nodes
git clone https://github.com/hyukudan/ComfyUI-MiniMax-H3-Prompt-Enhancer.git

Then restart ComfyUI. The pack itself is dependency-free - pure Python, GPL-3.0, Python 3.10+. What you do need to supply: a llama-server binary (grab it from a llama.cpp release) and a chat GGUF in ComfyUI/models/llm_gguf/. Nothing here downloads models for you - both the security upside and the one extra step.

Where people get burned

• "(llama-server not found)" in the dropdown - the binary isn't on PATH or in a ComfyUI-managed runtime dir. Set llama_server_path explicitly or install llama.cpp.
• Server won't start - check that context_size and gpu_layers fit your hardware. A 131072-token context on an 8GB card is a no-go; drop the context and offload less.
• Dialogue came out in English when you wrote Spanish - dialogue_language defaulted to auto and the detector misread the source. Set it explicitly.
• A one-off legal note if you're in the US/EU/UK/Korea: H3's own weights are geofenced out of your region by the MiniMax community license. That's about H3, not this node - but it's why some run the enhancer locally and still hit the model via API.