MOSS-Image SFT DeBand (int8)

MOSS-Image SFT with its vertical color stripes ("banding") fixed, as single files you load like any other model: no LoRAs, no custom nodes. Each file is 5.4 GB int8 (the same format as Comfy-Org's int8_convrot files) and loads with the stock Load Diffusion Model node.

These are the merged, ready-to-run versions of MATLOWAI/MOSS-Image-SFT-DeBand. That repo has the LoRAs, the full results and the write-up.

Pick one

How to use

1. Put the file in ComfyUI/models/diffusion_models/.
2. You also need the MOSS-Image text encoder and VAE from Comfy-Org/MOSS-Image:
   • qwen3.5_9b_int8_convrot.safetensors → models/text_encoders
   • flux2-vae.safetensors → models/vae
3. Drag workflows/moss_deband_int8.json into ComfyUI. Pick your file in Model and the same Version at the top; the version sets the steps.

Already have a MOSS-Image workflow? Just swap the model for one of these and use 25 steps (10 for Fast). Settings used here: ER-SDE sampler, CFG 3, Temporal Score Rescaling and skip-layer guidance (layers 22, 24).

Want 4–8 steps? The Turbo presets need the rapid-moss custom nodes, so they can't be merged into one file, and they only work near 1 MP (bigger sizes tile). For big images, DeBand Fast here is the quick option. See the rapid-moss workflow in the main repo.

Examples

Same prompt and seed (11), 768×1024. The color view removes brightness and boosts color 5×, which makes the stripes easy to see. Banding: about 1.0 = clean.

Fjord sky Goose at the checkout Oil still life Perfume bottle

How close is int8 to full weights?

Each cell is banding · similarity to the same setup in bf16 (SSIM; plain int8 SFT is compared with bf16 SFT, DeBand with the bf16 Quality preset, and so on).

• DeBand and DeBand+ match bf16 closely: about as close as Comfy-Org's int8 SFT is to bf16 SFT. The banding scores match the bf16 presets (1.39 and 1.09 on average).
• DeBand Fast drifts more from bf16 Fast: at 10 steps small weight differences move details the way a nearby seed would. The quality and the banding fix are the same.
• Wide images behave the same as bf16 (tested on the fjord sky): DeBand+ stays clean at 1536 and 2048 px wide; plain DeBand still shows faint wavy stripes on the sky there.

How they were made

• DeBand / DeBand+: Comfy-Org's int8 SFT with layers 37–41 (DeBand) or 32–41 (DeBand+) copied from Comfy-Org's int8 Base. No requantization; every tensor comes straight from their files. The full-weight equivalent is the banding-fix LoRAs.
• DeBand Fast: bf16 SFT with layers 37–41 from Base, plus lodestones' 8-step DMD LoRA (lora_latest) at 0.35, quantized to the same int8 format (per-channel int8, ConvRot 256). This recipe reproduces Comfy-Org's int8 SFT file to within one int8 step on fewer than 0.001% of weights.

SHA-256:

• DeBand 25082f19aaa11a4ebb3309080b5dd8bce59f0557bfd542976a7d6287e4869634
• DeBand+ da4e363d8206f41ffacf95787d7e0a67414dc3adb0636f27de338071d85f9cb5
• DeBand Fast ae5470cc8afb6a1b632d19ad28156c47ce9631fdceb69f3239ccceccd03a1e33

Limitations

• Small test set: five prompts at one seed, two prompts for wide images. Seeds swing the banding score a lot.
• Needs a ComfyUI recent enough to load Comfy-Org's int8_convrot MOSS-Image files.
• DeBand Fast trades some SFT detail for speed, like the Fast preset it comes from.

Credits

• OpenMOSS for MOSS-Image Base and SFT (Apache-2.0), and Comfy-Org for the int8 ConvRot files these are built from.
• lodestones for the 8-step DMD LoRA baked into DeBand Fast (Apache-2.0).
• Kijai for the SFT-difference LoRAs and the note that started the banding investigation.