Back to Model
M

MOSS-Image 4-Step Turbo

Model Oct 17, 2026

Akhaliq's community four-step distillation of MOSS-Image-SFT, with downloadable adapters, a merged transformer and a public Gradio demo.

Visit primary link 2 sources attached
Author-generated three-panel comparison of a concert-poster prompt with large English lettering and smaller bilingual details.
Generated samples from akhaliq's evaluation linked in the English four-step model card. Full comparison frame preserved; these are author-published outputs, not real event posters, an AITrending reproduction or a broad benchmark. Source: akhaliq / MOSS-Image Turbo evaluation
multimodalself-hosted

What happened lately

MOSS-Image 4-Step Turbo timeline

MOSS-Image 4-Step Turbo builds on an eight-step MOSS-Image-SFT distillation and publishes a four-step sampler, FP32 adapters and a merged transformer. The author recommends keeping the adapters in FP32 alongside the base BF16 transformer, and warns that a BF16 merge degrades fine text.

The Apache-2.0 card reports approximately 11 seconds per 1024-square image on an A100 and provides a small qualitative comparison set. These are author-reported results. Its canonical repository and demo were created October 17, while an earlier mirror preserves prior distribution; those dates do not establish the first release of the underlying work.