MiniMax-H3 Prompt Writer (T2VA/I2VA/FL2VA/L2VA)

Here's the honest trade this pack makes you choose between: the 27B rewriter gets the format from a trained LoRA, and it costs ~16 GB of VRAM plus a big download. This node gets it from MiniMax's own prompt-writing guide - the same document their documentation points at - dropped into the system prompt of an ordinary instruction-following GGUF. No LoRA, no 27B, no 52 GB anything.

The result: a 2.6 GB Qwen3.5-4B fills all three H3 fields in about 20 seconds on an 8 GB card. That's the whole pitch. The trade is real and worth stating plainly - the LoRA is the format, a 27B trained until H3 output came out of it, so its prose is denser and its formatting more reliable. Here, ~4,000 tokens of instructions have to hold the format together, and a small model obeys instructions less strictly than a trained distribution reproduces itself. Expect this to run at all on hardware the LoRA cannot touch.

How it works

The guide is fetched from MiniMaxAI/MiniMax-H3 on first use (~16 KB) and cached in ComfyUI/user/minimax_h3_rewriter/guides/. It lands in the system prompt, your prompt becomes the user turn, and the model writes in the guide's shape. Outputs are identical in name and order to the 27B rewriter's - rewritten_prompt, integrated_multimodal_description, overall_soundscape, non_diegetic_music - so the two are drop-in interchangeable in a saved workflow. Nothing here has to be Qwen, and nothing has to be installed: without llama-cpp-python the node runs the official llama.cpp binaries.

The inputs that matter

• prompt - the short idea.
• model - any GGUF with a chat template. Suggested starting points: Qwen3.5-4B on an 8 GB card, Qwen3.5-9B for the best writing per gigabyte, the abliterated 9B "Uncensored" build if the stock model declines a scene, or Gemma 3 12B / Mistral Small 24B when you have the VRAM.
• task - T2VA, I2VA, FL2VA or L2VA. Everything but T2VA also emits the alignment instruction line H3 requires as the very first line, with the duration already substituted to two decimals.
• reference_material - this node reads text, not pixels. For I2VA, FL2VA and L2VA you must describe what the reference frames show - by hand, or from a MiniMax-H3 Reference Caption node upstream. Without it the model invents a first frame that has nothing to do with your image. This is the trap in this node.
• greedy - keep on. Small models drift out of the format the moment they sample.
• duration - 4–15 s; the slider drives shot count and pacing. Keep it equal to your H3 settings.

n_ctx is handled for you: the base guide needs ~9,200 tokens of context and the full-reference guide ~12,300, against an 8192 default that suits the LoRA's shorter system prompt - so the node raises it automatically rather than letting llama.cpp truncate the front of the prompt (the guide and the output contract), which would silently break the format.

Install

Standard pack install - ComfyUI Manager (search "MiniMax-H3"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/pytraveler/MiniMax-H3-Prompt-Rewriter-ComfyUI

The writer nodes run the llama.cpp binaries, so the Python requirements are only needed for the Transformers-based rewriter nodes. First run fetches the guide and your chosen model.

Where people get burned

• Missing field. If a section comes back empty the node still returns everything it got and names the absent field on the node itself. Fix: lower the temperature, or move up a size. A 4B holds the format; a 9B holds it and writes better prose.
• Wrong first frame. You picked I2VA but didn't describe the frame in reference_material. It's text-only here - the model cannot see your image, so describe or caption it first.
• Format wobble. It's instructions being followed, not a trained distribution. Keep greedy on and skim the output before it goes into H3. That's the price of running on hardware the LoRA can't touch - and for most people it's a very good price.