8-step DMD distillation of MOSS-Image-SFT posted on Hugging Face
Original title: lodestones/MOSS-Image-SFT-8-steps-dmd — new model trending #29 on Hugging Face
The work-in-progress LoRA reduces a 50-step, CFG-3 image model to eight evaluations without classifier-free guidance.
A community checkpoint distills OpenMOSS-Team/MOSS-Image-SFT from 50 steps with CFG 3 to eight network evaluations and no CFG, using data-space DMD. The release, trending on Hugging Face, provides LoRA adapters for ComfyUI and warns that the transformer delta is smaller than bf16 rounding. Checkpoint update 2600 (EMA) uses float32 weights trained on compact-JSON scene_graph captions. The card marks the work as in progress.
- Recommended LoRA strengths are 1.0 for lora.safetensors and 0.7 for lora_latest.safetensors.
- LoRA rank and alpha are 128 on listed projection and feed-forward modules.
- Authors say bf16 casting discards most of the distillation delta.
- The sampler uses eight step codes and can run fewer prefix steps.
- Organizations
- OpenMOSS-Team
From the original article
1 min read at huggingface.cohttps://huggingface. co/spaces/hugging-apps/moss-image-8-step-dmd Work-in-progress distillation of [OpenMOSS-Team/MOSS-Image-SFT](https://huggingface. co/OpenMOSS-Team/MOSS-Image-SFT) (50 steps, CFG 3) to 8 network evaluations, no CFG.
By lodestones · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.