Original title: MOSS-Image-heal
The author states that the original MOSS-Image-Base has a few channels with extremely large values in its residual stream, approaching the fp16 limit; this full fine-tuning trains only the transformer to suppress these extreme values while keeping the output close to the original. The repository provides diffusers format and ComfyUI single files (bf16, fp16, int8 ConvRot), with the principles documented in the author's KohakuMossTrainer documentation.
Why it matters Provides a numerically more stable base model for those running MOSS in half-precision or continuing training on it, with image outputs close to the original.
Hugs 35 · Hug Score Increase 34 · Hugs Downloads 0
License: Commercial use allowed · ComfyUI: Mentioned in model card · VRAM: ≈13 GB estimated
Reasons
- Apache-2.0 license
- The model card includes comparison images with identical prompts for the original and fine-tuned versions, along with all statistical reports
Usage
- For ComfyUI, replace the MOSS Base transformer with the single file from comfyui/; the author's sample settings are 16 steps, CFG 3, shift 5
Risks
- This is a low-level improvement aimed at training and low-precision inference; image style will not change noticeably; differences in image quality should be judged based on the author's comparison images