MiniMax H3 Prompt Enhancer (Cloud / Local GGUF)

MiniMax H3 is the open-weight video model that launched to a chorus of "unbelievably good," and its one grumpy requirement is that you prompt it the way its docs insist. Not "a woman walks through a neon street." Proper H3 form - [Shot N], strict timestamps, a speaker contract, a soundscape field the model actually expects by name. Hand it sloppy prose and you'll get a fine-looking clip that quietly ignores half of what you meant.

This node is the translation layer. You type the rough idea, it pays an LLM to rewrite it into official H3 shape, and you get one enhanced_prompt STRING to feed into your H3 generation node. It's a prompt enhancer, not a generator - it operates on text, upstream of the sampler, which is the entire point.

How it works

The task_type dropdown picks the H3 format you're aiming at: T2VA (text to audio-video), I2VA (first-frame), FL2VA (first + last frame), L2VA (last frame only), or Ref2VA (reference images/videos). The official core h3-prompt-writing skill is always baked in, pinned to a fixed upstream commit, so details like (S1) speaker tags, <scenetrans> for cross-cut dialogue and <cutoff> for end-of-clip cutoffs come out right without you knowing the syntax. shot_count goes AUTO or a fixed 1–20; a fixed number becomes a hard "write exactly this many [Shot N]" instruction, though the model is allowed to miss. rewrite_mode (strict / balanced / creative) maps to temperatures 0.2 / 0.7 / 1.2 - creative is where the writing gets visibly punchier.

Beyond the core there's optional flavor: 8 official scene presets (product ad, 3D short, paper-collage explainer, the official music-video subtitle generator, and more) and a library of 190 "case templates" that transplant structure and pacing from real published clips without copying their characters or plot. The two dropdowns are easy to ignore; the core does the job alone.

The inputs that actually matter

On a first run you set four things: prompt (the only required field), task_type, shot_count or duration_seconds, and rewrite_mode. output_language defaults to 中文 - flip it to English if that's how you prompt. If you picked anything but T2VA, wire the matching first_frame, last_frame, reference_images (up to 9) or reference_videos (up to 3) sockets, and the LLM genuinely looks at them. reference_context and constraints are folded away for good reason: they exist for facts the pixels can't tell it (who's who, what must not change). Leave them alone until you hit a case they solve.

There are four channels behind api_mode:

• 贞贞平价小屋 (default): https://api.seedance.nz, fixed vision model bytedance/doubao-seed-evolving.
• 贞贞的 AI 工坊: https://ai.t8star.org/v1/chat/completions, default gemini-3.5-flash.
• OpenAI compatible: your own base URL + model ID, for whichever provider you already pay.
• 本地 GGUF: ComfyUI/models/LLM, no key at all, Qwen3.8-27B recommended.

A note on the author: T8mars runs the first two gateways, so the cloud half is partly a funnel to paid endpoints (there are affiliate links in the README). It's a real, well-tested product though, and the local GGUF path sidesteps the funnel entirely.

Install

ComfyUI Manager → search "MiniMax H3 / Seedance 2.0 / Music 3 Prompt Enhancer (T8)", or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-prompt-enhancer-T8.git

Restart ComfyUI, hard-refresh the browser with Ctrl+F5 if the menu doesn't update. Cloud mode needs no extra Python packages. The local path is the heavy one: run python install_local_qwen.py inside the pack directory to pull the ~15.9 GiB Qwen GGUF, the ~0.9 GiB vision projector, and llama.cpp. Plan for 24 GB VRAM; 16 GB works via --fit offloading but is slow.

Where people get burned

• The key is for the enhancer LLM, not for video generation. This key bills the prompt-rewrite requests. Your downstream H3 generation API key is a different thing entirely.
• Clicking "save to workflow" embeds the key in the workflow JSON. Clear it before you share anything.
• Cloud videos cap at 50 MB (MP4/AVI/MOV/MKV); Ref2VA wants 2–15s clips, 12 total pieces of media max.
• Local mode samples your video at ~2 fps and reads only the visuals - it will not, and does not claim to, hear the audio track.
• Gateway 5xx and Cloudflare hiccups retry automatically up to three times; a 401 or empty balance won't. That's by design, not a stuck node.