MOSS-Image Base vs SFT

Same architecture, same text encoder and VAE, same settings. SFT is Base fine-tuned on curated data. Use SFT to make images, Base to train on.

The official comparison from the model card: SFT groups the text more clearly and lays out the whiteboard more neatly.

	Base	SFT
Trained	Pre-training only, no post-training or CFG distillation	Supervised fine-tune of Base
Best for	LoRAs, fine-tunes, research, A/B tests	Generating images
Text and layout	Good	Clearer grouping, tidier layouts
Known issue	Needs good settings: higher CFG for short prompts	Vertical stripes on smooth areas; can look overblown
Turbo LoRAs	None yet	8-step and 4-step
File (ComfyUI)	moss_image_base_int8_convrot / bf16	moss_image_sft_int8_convrot / bf16

Worth knowing: all 17 images in the official example set, the ones on our prompts page, were made with Base, according to the parameters files saved with them. Base is not just for training.

If you train on Base, the user guide asks you to keep the text encoder frozen, and gives the captioning prompt the team used (Gemma 4 or Qwen 3.5 27B / 35B-A3B with thinking on).

Fixing the SFT stripes

SFT often draws faint, evenly spaced vertical colour stripes across skies, fog, backdrops and walls, most visibly on wide images of 1536 px and up. Base does not. Three community fixes:

• DeBand LoRA: moss_image_sft_banding_fix_L37-41_bf16.safetensors in a LoraLoaderModelOnly at 1.0. It swaps SFT's last five layers back to Base weights. If skies still stripe, add the L32-36 “extra” LoRA at 1.0 as well.
• DeBand INT8 files: the same fix merged into 5.41 GB files you load like the normal model. DeBand (25 steps), DeBand+ for wide images, DeBand Fast (10 steps).
• Kijai's SFT LoRA for blocks 21–49 at strength −0.5: removes the stripes and tones down the overblown SFT look. ComfyUI's own template notes point to it.

Or generate with Base. The DeBand author's write-up traces the stripes to SFT's last layers, which is why restoring those layers from Base removes them.