Gemma4-for-MiniMax-H3-Prompt-Extend

This is a prompt enhancement model fine-tuned by Wuli-art using Gemma4-12B-IT, specifically trained for MiniMax-H3 audio and video generation: the FL2VA format covers text-only, first-frame, last-frame, and first-and-last-frame generation, while the Ref2VA format covers reference-image generation, with both sharing the same set of weights. The repository provides a standard BF16 version and FP8 / Int8 quantized versions, along with two separate ComfyUI workflows; the author also created a comparison page that plays side-by-side 45 videos across 9 cases, comparing no expansion, the original Gemma4, H3 official expansion, lightx2v rewriting, and this model.

This is a prompt enhancement model fine-tuned by Wuli-art using Gemma4-12B-IT, specifically trained for MiniMax-H3 audio and video generation: the FL2VA format covers text-only, first-frame, last-frame, and first-and-last-frame generation, while the Ref2VA format covers reference-image generation, with both sharing the same set of weights. The repository provides a standard BF16 version and FP8 / Int8 quantized versions, along with two separate ComfyUI workflows; the author also created a comparison page that plays side-by-side 45 videos across 9 cases, comparing no expansion, the original Gemma4, H3 official expansion, lightx2v rewriting, and this model.

Why It Matters It organizes a single requirement into scene, action, camera movement, and sound descriptions. Both first/last-frame and reference-image tasks share the same weights, allowing it to be directly chained before H3 in ComfyUI.

Hugging Face Likes 5 · Hugging Face Trending Score 5 · Hugging Face Downloads 0

License: Commercial Use Allowed · ComfyUI: Mentioned in Model Card · VRAM: ≈29 GB (estimated)

Regional License: The MiniMax H3 open-weight license excludes the EU, UK, South Korea, and the US (see Excluded Territories in the LICENSE); users in these four regions must obtain separate authorization from MiniMax.

Reasons

  • Apache-2.0 license, with downloads also available via ModelScope
  • The comparison page allows synchronized playback and expansion of both original and enhanced prompts

Usage

  • Download the model files and place them in ComfyUI. Drag minimax_h3_gemma4_pe_fl2va.json or minimax_h3_gemma4_pe_ref2va.json onto the canvas, select the standard or quantized version in the PE loader node, and after running, connect the enhanced result to the H3 generation workflow

Risks

  • The BF16 version of the 12B model requires approximately 29 GB of VRAM based on parameter count estimates; choose the FP8 / Int8 version if you have limited VRAM. The workflow also requires installing the custom nodes mentioned in the model card
  • The enhanced results must be used with MiniMax-H3, and the H3 license does not apply to the EU, UK, South Korea, and the US

Open

抱脸 image-to-video 新模型

Explore the latest AI Radar updates