Guide
Should I use MOSS-Image Base, SFT or Turbo?
Use SFT to make images; it is fine-tuned on curated data and gives clearer text grouping and tidier layouts. Consider Base for LoRA training or fine-tuning. Turbo builds (4 or 8 steps) are community few-step versions based on Base or SFT. They target faster generation; quality depends on the build and scene.
MOSS-Image Base
The pretrained model. The official guide says it has had no post-training, to preserve training information for further LoRA training. It also needs suitable settings: 50+ steps, CFG 3–6 (raise CFG for short prompts), shift 5. If you fine-tune it, keep the text encoder frozen.
MOSS-Image SFT
A supervised fine-tune of Base on curated data, with the same architecture and pipeline. In the official comparison, SFT produces clearer text grouping and a more organized layout. Default settings: 50 steps, CFG 3, shift 5, Euler. MOSS-Image AI uses SFT.
Turbo and few-step versions (community)
| Version | Steps | CFG | How to use |
|---|---|---|---|
| lodestones/MOSS-Image-SFT-8-steps-dmd | 8 | none | LoRA at strength 1; author calls it work in progress |
| akhaliq/MOSS-Image-SFT-4-steps | 4 | 1 | One LoRA on the stock flow-matching Euler sampler (v2.0, Oct 18) |
| lodestones/MOSS-Image-Base-8-steps-dmd | 8 | none | Same method, applied to Base |
The official team says few-step distillation such as DMD is in progress.
Which one should you pick?
- Making posters, menus or covers: SFT.
- Training a style or character LoRA: Base.
- Iterating quickly: try a 4-step or 8-step build, then render finals with SFT.
Fewer steps mainly reduce generation time; they do not imply smaller weights or lower peak VRAM. Check the VRAM guide before choosing hardware.
For official models and settings, see the OpenMOSS documentation and user guide.
Questions
Do Base and SFT use the same files otherwise?
Yes. Same architecture, text encoder and VAE; only the diffusion model differs.
Can I use SFT LoRAs on Base?
LoRAs are trained against one model. Use them on the model they were trained for.
Is Turbo quality as good?
Few-step distillations usually lose some detail. Compare on text-heavy prompts before switching.
Which version does the official demo use?
The official Gradio app lets you choose the model version.
Will there be an official Turbo?
The team says few-step distillation is in progress, with no date.
Turn your idea into an image
Describe your idea, generate an image, then refine the text or layout if needed.