Reducing a 50-step image-generation process to eight evaluations is not just a speed tweak: it changes how the model uses its own intermediate predictions. This experimental distillation aims to retain the parent model’s image-generation behavior while removing classifier-free guidance and most repeated transformer calls.
What Sets It Apart
- Data-space DMD: Each iteration feeds the generator’s clean estimate back alongside fixed noise, rather than simply replaying the original scheduler. This is the central change behind the short sampling trajectory.
- No CFG at inference: The target is eight network evaluations without the usual guided and unguided forward-pass pair, reducing inference work compared with the parent model’s default setup.
- Scene-graph-oriented training: Training uses compact JSON
scene_graphcaptions, connecting the distillation to structured descriptions of image content and layout. - An early checkpoint, not a finished replacement: The published EMA checkpoint is update 2600 of a planned 12000. Treat quality and compatibility as experimental rather than assuming parity with the 50-step source model.
Who It’s For
A good fit for MOSS-Image-SFT users who want to experiment with faster generation or structured layout prompts, especially in ComfyUI. Look elsewhere if you need a stable, fully validated drop-in replacement: this is work in progress, and its distilled behavior depends on the supplied adapter and sampler setup.
