SFT Destripe v1.0

MOSS-Image SFT paints faint, evenly spaced vertical colour stripes across smooth areas such as skies, fog, water, walls and studio backdrops. This 1.9 MB LoRA removes them and keeps the SFT look.

Builds on MATLOWAI's DeBand investigation His block-swap analysis found the layers that carry the stripes, and his bandviz tool is the banding score used here. Go give it a look.

How to use

1. Load MOSS-Image SFT moss_image_sft_bf16 or moss_image_sft_int8_convrot). Not Base.

2. Add LoraLoaderModelOnly with this LoRA.

3. Set the strength:

- 1.0 at 768 px

- 1.5 at 1024 px

4. Sample as usual. No trigger words.

Results

Blind test on 12 images that weren't used to build the LoRA. The banding score is from bandviz: lower is cleaner, and about 16 or below looks clean.

At 1024 px

- SFT: 24.0 (stripes in 11 of 12 images)

- SFT + this LoRA @ 1.5: 5.9 (stripes in 1 of 12)

- Base model: 4.0

At 768 px

- SFT: 9.8

- SFT + this LoRA @ 1.0: 4.2

- Base model: 4.2

Similarity to plain SFT stays above 99% (SSIM 0.993–0.996). The image barely changes except for the stripes. For comparison, reverting whole layers to Base scores 8.4 at 1024 px, with 98.3% similarity, and takes a 1.13 GB LoRA. This one is 1.9 MB.

Independent comparison

MATLOWAI has put this LoRA side by side with SFT, his own DeBand fixes and several other candidates. Every fix gets the same prompts and seeds, and pan and zoom move together across columns: MOSS-DeBand Comparison

His results match the ones above. Of the fixes that remove stripes, this one stays closest to SFT: face similarity to SFT is 0.94, against 0.83–0.85 for the other fixes. It does that by changing about a thousandth of what SFT changed. The heavier fixes remove more stripe on flat colours but move the image further from SFT. If stripe removal matters more than keeping the SFT look, those are worth trying.

The Edit Anatomy page shows where each fix changes the model, and how much stripe each removes against how much of SFT it undoes: Edit Anatomy

Seeing the stripes

The bands are a colour shift, not brightness, so they're hard to see until you know where to look. Every gallery image shows the render with a colour-boosted view underneath, where the stripes are obvious.

How it was made

It's extracted, not trained:

1. SFT's change from Base is split into directions, layer by layer.

2. A gradient probe measures which directions feed the stripes while the image forms.

3. The 64 strongest are subtracted from SFT, and everything else SFT changed is left alone.

Full write-up and numbers: https://huggingface.co/nynxz/MOSS-Image

Notes

- Tested with ER-SDE, 25 steps, CFG 2.5 (the official settings). Other samplers are untested.

- Load LoRA (Bypass) drops the q/k/v parts of this LoRA on the fused Comfy-Org checkpoints. Use normal LoRA loading, or the split q/k/v checkpoints from the HF repo.

- Banding varies a lot from seed to seed, and not every SFT image shows it.

Credits: MATLOWAI for the original banding investigation, the layer analysis and bandviz. OpenMOSS-Team for MOSS-Image.