Original title: chroma-h3
lodestones fine-tuned the MiniMax H3 video DiT into an image generation (stills) model by training on single latent frames using text-image pairs. The repository provides a merged full pipeline, a fine-tuned LoRA that can be added to the original H3, and a few-step version distilled from it: specifically, the lora_vspace_dmd8 file includes both the fine-tuning and few-step capabilities, with the author stating that sampling can be done using Euler, CFG 1, 4, or 8 steps.
Why it matters H3 users can generate still images without changing the base model, and the few-step LoRA can be directly added to the original H3 to increase speed.
Hugging Face likes 50 · Hugging Face trending score 49 · Hugging Face downloads 86
License: Conditional commercial use · ComfyUI: Likely compatible · VRAM: Not specified
Regional license: The MiniMax H3 open-weight license excludes the EU, UK, South Korea, and the US (Excluded Territories in the LICENSE); users in these four regions must obtain separate authorization from MiniMax.
Reasons
- The author clearly specifies in the model card which transformer each LoRA should be applied to, along with strength and sampling settings
- Includes LoRA files in ComfyUI format
Usage
- Load lora/chroma-h3_vspace_dmd8-comfyui.safetensors on the original H3 with strength 1, using Euler, CFG 1, 4, or 8 steps; for still image generation, use shift 3
- For fine-tuning only without few-step: load lora/chroma-h3-comfyui.safetensors with strength 1
Risks
- Inherits the MiniMax H3 community license, which excludes the EU, UK, South Korea, and the US; users in these four regions must obtain separate authorization
- The few-step LoRA cannot be stacked on the merged transformer already included in this repository, as specifically warned in the model card; the author has not provided evaluations beyond comparison images