MOSS-Image Base vs SFT - which checkpoint should you use?

OpenMOSS released MOSS-Image as two checkpoints on October 16, 2026: MOSS-Image-Base and MOSS-Image-SFT. Both are the same 5B diffusion transformer with the same text encoder and decoder, and both load with the same pipeline. The difference is what happened after pretraining.

Base: the raw pretrained model

Base is the model as pretraining left it. OpenMOSS says it deliberately did not post-train or distil its inference behaviour, so that the full training signal stays intact for people who build on it.

That has two consequences:

• It is the right starting point for training. LoRAs and full fine-tunes learn more cleanly from an untouched base. The user guide adds one rule: keep the text encoder frozen when you post-train.
• It needs good settings to look its best. The user guide recommends at least 50 steps, guidance between 3 and 6 (higher for short prompts), shift 5 and skip-layer guidance on layer 22. With poor settings, Base images can look flat or lose detail.

SFT: tuned for everyday images

SFT is Base after supervised fine-tuning on curated data. The model card's side-by-side example, a whiteboard covered in handwritten notes, shows SFT grouping the text more clearly and organising the board more tidily than Base.

For most people generating images, that is what you want: more consistent composition and cleaner text from ordinary prompts, at the same default settings.

Side by side

	Base	SFT
Architecture	5B DiT, Qwen3.5-9B encoder, FLUX.2 small VAE	Same
Training	Pretraining only	Pretraining plus supervised fine-tuning
Best for	LoRA and fine-tune training, research	Generating images
Out-of-the-box results	Needs careful settings	More consistent
Licence	Apache 2.0	Apache 2.0

Which one runs on MOSS Image AI?

We run MOSS-Image through hosted inference providers that serve the generation checkpoint and choose sampling settings for it, so you don't need to tune steps or guidance. If you want to train your own LoRA, download Base from Hugging Face or ModelScope and see our ComfyUI guide for the local setup.

Bottom line

• Making images: SFT.
• Training a LoRA or a custom fine-tune: Base, then use your adapter on Base or test it on SFT.
• Not sure: SFT. Try it in the studio with the free credits every new account gets.