OpenMOSS releases MOSS-Image base and SFT models

OpenMOSS published MOSS-Image-Base and its SFT sibling, open text-to-image models supporting prompts, tags, and JSON layouts.

OpenMOSS released MOSS-Image-Base, a base text-to-image model, and MOSS-Image-SFT, a supervised fine-tune of that base. Both accept natural-language prompts, tag sequences, and JSON layouts to control text placement and object arrangement. The Base card documents Diffusers inference with a custom pipeline, flow-matching Euler sampling at 50 steps and CFG 3, bfloat16 for the transformer and text encoder, and float32 for a FLUX.2 small VAE decoder; the SFT report says that model keeps the same architecture and inference pipeline with the same sampler defaults. A technical report and ComfyUI plugin are available, and the SFT release also lists weights on Hugging Face and ModelScope plus a demo Space. Sources disagree on ranking: one report places MOSS-Image-Base at Hugging Face trending #30, while the other attributes the #30 slot to MOSS-Image-SFT.

• OpenMOSS released MOSS-Image-Base and MOSS-Image-SFT, a supervised fine-tune of the base text-to-image model.
• Both accept natural-language prompts, tag sequences, and JSON layouts for text placement and object arrangement.
• Recommended sampling is flow-matching Euler with 50 steps and CFG 3, using a FLUX.2 small VAE decoder.
• The Base card documents Diffusers inference with bfloat16 for the transformer and text encoder and float32 for the VAE.
• A technical report and ComfyUI plugin are available; SFT weights are also on ModelScope, with a demo Space.
• Reports dated 2026-10-11 disagree on which model holds Hugging Face trending rank #30: Base or SFT.