Seedreamstyle is a photographic style LoRA (rank 16) trained by community author Nebsh for MiniMax H3 Image. It targets portrait, conversational, and character scenes across different ages and real-world lighting conditions, aiming to bring the visual characteristics of Seedream 5.5—specifically its contrast, lighting, and grain—to H3. The 2K version was trained on 43 images, then continued training on 1,134 images up to 6000 steps. The author recommends the 6K version, and intermediate checkpoints from steps 2500–5500 are also provided. The author also tested it with H3 video for text-to-video and reference-image-to-video generation, including 4 sets of video comparisons with and without the LoRA, but notes that these are merely supplementary experiments.
Why it matters The model card includes twenty sets of four-image grids with the same seed (without LoRA, two training stages, and original Seedream images) and four H3 video comparisons with and without LoRA, making the style changes immediately visible.
Hugging Face likes 11 · Hugging Face trending score 11 · Hugging Face downloads 0
License: Conditional commercial use · ComfyUI: Mentioned in model card · VRAM: Not specified
Regional licensing: 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
- All twenty comparison sets include prompts, seeds, and parameters
- Includes all intermediate checkpoints and a SHA-256 manifest, allowing users to compare different training stages themselves
Usage
- In ComfyUI, load Seedreamstyle_H3_Resume_6K.safetensors using the LoRA Loader (Model Only) at strength 1.0, and start prompts with "seedreamstyle."; the author's comparisons used the int8 pruned version of H3 Image, LCM sampling with 6 steps, CFG 1, and simple scheduling, along with KJ's MiniMaxH3PromptWeight node
Risks
- Follows the MiniMax H3 Image community license, which does not apply to the EU, UK, South Korea, and the US
- Of the 1,134 training images, only about 150 are documented as coming from two donors; the sources of the rest are not specified
- The output resolution of H3 differs from the Seedream reference images, so comparisons reflect visual appearance only, not detail comparisons under identical conditions