Model guide · checked Oct 22, 2026
MOSS-Image SFT vs Base: which one to use
Both are 5B text-to-image models from OpenMOSS with the same architecture, text encoder and VAE, and their files are the same size. SFT is Base fine-tuned on curated data. Generate with SFT, train LoRAs on Base, and know about SFT's one common flaw: faint vertical color stripes.
The short answer
- Making images: start with SFT. In OpenMOSS's own side-by-side, SFT groups text more clearly and lays the scene out more neatly.
- Training a LoRA or fine-tuning: use Base. OpenMOSS left it without post-training so it keeps all of its training and is easier to train on.
- Thin vertical color stripes in a sky or a plain backdrop are an SFT artifact. Add one of the fix LoRAs below, or switch to Base, which doesn't draw them.
Side by side
| SFT | Base | |
|---|---|---|
| What it is | Base fine-tuned on curated data (supervised fine-tuning), with the same architecture and pipeline | The base text-to-image model, with no post-training |
| Use it for | Generating images | LoRA training and fine-tuning |
| Text and layout | Clearer text grouping and a tidier layout in OpenMOSS's example | Looser grouping in the same example |
| Vertical color stripes | Often, on smooth areas such as skies, fog, walls and backdrops | None; the community fixes work by restoring Base weights |
| 4- to 10-step speed-ups | The community 8-step and 4-step adapters are trained on SFT | None in the official README's community list |
| ComfyUI templates | MOSS-Image: Text to Image · MOSS-Image: Layout-Guided Text to Image | MOSS-Image Base: Text to Image · MOSS-Image Base: Layout-Guided Text to Image |
| Diffusion model file |
|
|
| Text encoder and VAE | The same Qwen3.5-9B text encoder and FLUX.2 VAE files | The same Qwen3.5-9B text encoder and FLUX.2 VAE files |
| License | Apache-2.0 | Apache-2.0 |
Settings: the same for both
Both models run in the same workflow; only the diffusion model file changes. The four ComfyUI templates share one set of settings. OpenMOSS's user guide, written for Base, asks for more steps: Base has not been post-trained for settings such as CFG, so it needs suitable ones to look its best.
| Setting | ComfyUI templates | OpenMOSS guide |
|---|---|---|
| Steps | 25 | 50 or more |
| CFG | 2.5 | 3; up to 6 for short prompts |
| Sampler | er_sde · simple | Euler |
| Shift | 5 (the model default) | 5 |
| Skip-layer guidance | Layer 22, first half of the steps | Layer 22; 24 and 26 also help |
| Image size | 1024×1024 | 1 to 2 megapixels, sides in multiples of 16 |
Fixing SFT's vertical stripes
The stripes are a shift in color, not brightness, so they are easy to miss until you look at a sky or a plain wall: about 21 px apart at 1024 px wide and 16 px at 768 px. Each fix below loads with the core LoraLoaderModelOnly node after the SFT model, except the merged int8 files, which replace the model file.
- Destripe LoRA (nynxz, 1.9 MB): the smallest change to the picture. Strength 1.0 at 768 px wide, 1.5 at 1024 px.
- DeBand (MATLOWAI): at strength 1.0 it turns SFT's last five layers (37–41) back into Base. For images 1536 px wide and up or big smooth skies, stack the extra fix for layers 32–36.
- DeBand int8: the same fix merged into ready-to-load 5.4 GB models, used in place of moss_image_sft_int8_convrot.safetensors. No LoRA or custom node.
Prefer the opposite? Kijai published the full SFT difference as a LoRA, so you can put the SFT look on Base at any strength; at full strength it brings the stripes back too.
Training
Two trainers listed in OpenMOSS's README support full fine-tuning and LoRA for both models: KohakuMossTrainer, which trains adapters in as little as 6 GB of VRAM, and DiffSynth-Studio. OpenMOSS released Base for training and says to keep the text encoder frozen.
Download
Pick the model and your GPU memory in the file picker to get the exact files and download commands, or start from a ready-made workflow.
Sources: OpenMOSS/MOSS-Image README · OpenMOSS-Team/MOSS-Image-SFT · OpenMOSS-Team/MOSS-Image-Base · OpenMOSS user guide · Comfy-Org/workflow_templates · nynxz/MOSS-Image (Destripe) · MATLOWAI/MOSS-Image-SFT-DeBand · MATLOWAI/MOSS-Image-SFT-DeBand-int8 · Kijai/MOSS-image_experimental · KohakuMossTrainer · DiffSynth-Studio MOSS-Image docs