MOSS-Image SFT DeBand (INT8)

MOSS-Image SFT has a tendency to produce vertical color banding, particularly in skies, gradients, fog, and other areas of relatively uniform color.

DeBand addresses this by replacing a small number of transformer layers with their counterparts from MOSS-Image Base. This reduces the banding at the model level rather than trying to correct it through postprocessing.

The goal is to preserve the qualities of MOSS-Image SFT while getting rid of an annoying artifact.

This release includes three ready-to-use INT8 ConvRot checkpoints for ComfyUI. Each is approximately 5.4 GB, with no additional LoRAs required.

Versions

DeBand+

The stronger correction, replacing transformer layers 32–41 with their Base counterparts.

This produced the lowest average banding score in our tests and is particularly useful for landscapes, smooth gradients, and wider compositions where the artifact can become more apparent.

Recommended settings: 25 steps.

DeBand

A more conservative correction, replacing layers 37–41.

Retains more of the original SFT weights while still substantially reducing banding. A good starting point if you want results as close as possible to the original model.

Recommended settings: 25 steps.

DeBand Fast

The DeBand correction combined with lodestones' DMD acceleration, baked directly into the checkpoint.

Runs at 10 steps instead of 25, with some trade-off in fine detail.

Recommended settings: 10 steps.

Results

We tested the models across five prompts designed to expose banding in different situations, including skies, fog, and smooth gradients.

ModelMean banding score (lower is better)Original MOSS-Image SFT INT82.09DeBand INT81.36DeBand+ INT81.15DeBand Fast INT81.20

DeBand+ had the strongest average banding reduction, followed closely by Fast.

These are results from a small test set rather than an exhaustive benchmark. The Fast variant also uses a different sampling configuration, so its numbers aren't directly comparable in every respect.

Comparison images and additional details are available in the Hugging Face repository.

Installation

Download your preferred checkpoint and place it in:

ComfyUI/models/diffusion_models/

You'll also need the text encoder and VAE from Comfy-Org/MOSS-Image.

Text encoder:

qwen3.5_9b_int8_convrot.safetensors

Place in ComfyUI/models/text_encoders/

VAE:

flux2-vae.safetensors

Place in ComfyUI/models/vae/

These are diffusion-model-only checkpoints, so use ComfyUI's Load Diffusion Model node rather than the standard checkpoint loader.

ComfyUI Workflow

A ready-to-use workflow is available here:

Download the MOSS-Image DeBand INT8 workflow

Load it into ComfyUI, select your checkpoint, and adjust the sampling settings for the version you're using.

Sampling Settings

SettingDeBand / DeBand+FastSteps2510CFG33SamplerER-SDEER-SDETemporal Score RescalingEnabledEnabledSkip-Layer Guidance22, 2422, 24

These are the configurations used for the published comparisons. Other settings may work well too.

Technical Details

The banding appears to be associated with changes introduced during SFT rather than being inherent to the base model.

By comparing the two sets of weights, we found that replacing a relatively small group of later transformer layers could substantially reduce the artifact without reverting the entire model to its Base behavior.

DeBand replaces layers 37–41, while DeBand+ extends the replacement to layers 32–41.

The INT8 quality variants use the original Comfy-Org quantized weights for the selected layers, avoiding an additional quantization pass.

The Fast variant combines the DeBand correction with lodestones' DMD acceleration before quantization.

The approach is fairly simple, but the improvement is noticeable, especially in images where broad areas of similar color make the original artifact difficult to ignore.

Notes

These models require a recent version of ComfyUI with MOSS-Image and INT8 ConvRot support.

The text encoder and VAE are distributed separately. The 5.4 GB checkpoint size doesn't represent total VRAM requirements.

DeBand and DeBand+ generally preserve the original SFT's character, though replacing transformer layers can affect generation behavior beyond banding. Fast makes an additional quality/speed trade-off.

The correction isn't guaranteed to eliminate every instance of banding, particularly at unusual resolutions or with certain prompts.

Downloads and Research

INT8 checkpoints, workflows, comparisons, and checksums:

MOSS-Image-SFT-DeBand-int8 on Hugging Face

Original research, BF16 variants, and LoRA-based corrections:

MOSS-Image-SFT-DeBand on Hugging Face

Credits

• OpenMOSS for the original MOSS-Image models.

• Comfy-Org for the INT8 ConvRot releases and ComfyUI integration.

• lodestones for the DMD acceleration work used in the Fast variant.

• Kijai for the SFT-difference LoRAs and observations that helped motivate the banding investigation.

Released under Apache 2.0, subject to the applicable upstream component licenses.