MOSS-Image-SFT-8-steps-dmd: From 50 to 8 Steps: CFG-Free MOSS-Image Inference

A distillation checkpoint that compresses OpenMOSS-Team's MOSS-Image-SFT model into an 8-step network evaluation. It removes the original 50-step inference and CFG 3 setting, applying the Data-space DMD technique to increase generation speed. It operates by having the generator iteratively refine its own clean estimate from fixed noise.

The model is at the 5B parameter scale and consists of a transformer trained in fp32 precision and LoRA adapter files. Because the distillation effect is smaller than the bf16 rounding step, the transformer must be loaded and executed in fp32. Casting to bf16 largely eliminates the distillation effect, so for fast inference, it is recommended to apply the adapter without merging it into the bf16 transformer from the base repository.

In a ComfyUI environment, lora.safetensors can be loaded directly at strength 1.0 and lora_latest.safetensors at strength 0.7. It is released under the Apache-2.0 license, and pipelines can be constructed using the Diffusers library. It is also possible to use only some of the 8 steps, as each prefix is trained, supporting flexible inference.

It is suitable for developers who want to drastically reduce inference steps while maintaining image generation quality. It is particularly useful when stable results are desired without CFG, or when leveraging a lightweight LoRA in ComfyUI workflows. It is trained with compact-JSON scene_graph captions, specializing in text prompt understanding.