Wuli's Gemma-4 Prompt Extender for MiniMax H3 Runs in ComfyUI
Wuli-Art fine-tuned Gemma-4-12B into a MiniMax H3 prompt extender: one weight set for FL2VA and Ref2VA, fp8 at 13 GB, with two ComfyUI workflows.
Wuli-Art released Gemma4-for-MiniMax-H3-Prompt-Extend on 2026-11-16: a Gemma-4-12B fine-tune that rewrites a plain idea into the audiovisual prompt MiniMax H3 expects. It ships bf16, fp8 and int8 weights plus two ComfyUI workflows, and those graphs use core nodes only.
What it does
An H3 prompt is one block that has to carry the shot, the action over time, the camera move and the soundscape. Writing that by hand is the slow part. This model writes it from your sentence plus whatever frames or references you already loaded.
One weight set serves both input formats, so you do not need a different extender for first/last-frame work and for reference work.
| H3 input format | Covers | What the extender adds |
|---|---|---|
| FL2VA | T2VA, I2VA, L2VA, FL2VA | scene, action, camera movement, sound, frame constraints |
| Ref2VA | reference-conditioned generation | subject definitions, reference roles, what to keep |
How to run it
- Put
gemma4_12b_h3_pe_fp8_scaled.safetensors(13.0 GB) or the int8 file inComfyUI/models/text_encoders/. The bf16 file is 23.9 GB. - Drag the matching workflow onto the canvas: FL2VA or Ref2VA. They are alternatives, not two stages.
- Point the UNET, text encoder, VAE and LoRA loaders at your own H3 files, write the prompt, run. The output is text, which you feed to your normal H3 graph.
Version requirement: a ComfyUI build with TextGenerateV2 and Gemma-4 text-encoder support. v0.46.0 has both.
Downloads
| File | Size | Where it goes |
|---|---|---|
gemma4_12b_h3_pe_fp8_scaled.safetensors | 13.0 GB | models/text_encoders/ |
gemma4_12b_h3_pe_int8.safetensors | 13.0 GB | models/text_encoders/ |
gemma4_12b_h3_pe_bf16.safetensors | 23.9 GB | models/text_encoders/ |
Compared with the other prompt writers
The model card's demo runs five settings on the same inputs: no extension, plain Gemma-4-12B, MiniMax's official PE, lightx2v's rewriter, and this model. Case 04 starts from one 31-character Chinese line about two cranes fighting, plus a first frame.
![]() | ![]() | ![]() |
|---|---|---|
| No extension | MiniMax official PE | Gemma-4-12B extender |
Length is the visible difference. On this case the plain Gemma-4 baseline wrote 1,097 characters, lightx2v's rewriter 1,180, this model 2,102, and the official PE 4,296.
A 5-second 1024×768 sample generated from this extender's prompt.
Availability
- Weights: Hugging Face · ModelScope
- 45-clip five-way comparison: Wuli demo page
- Wuli-Art's earlier H3 release: MiniMax-H3-Turbo



Comments
Sign in with GitHub to join the discussion.