MOSS-Image SFT Destripe LoRA

MOSS-Image SFT Destripe LoRA

Removes faint vertical color stripes from MOSS-Image SFT pictures in ComfyUI with a 1.9 MB LoRA, and adds checkpoints with q, k and v split apart. For ComfyUI users who generate with MOSS-Image.

Made by nynxz on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This repository holds two add-ons for running MOSS-Image in ComfyUI. One is a tiny LoRA that cleans up pale vertical bands of color, which the SFT model leaves across flat regions like sky, mist, walls and photo backdrops.

The other is two checkpoints that match Comfy-Org's weights, except that the attention q, k and v sit in separate tensors.

Who it suits: It suits people who generate images with MOSS-Image SFT in ComfyUI and see color stripes in flat areas. The bands change hue rather than brightness, so you may only notice them once you strengthen the color.

Industry | Who uses it | How they use it | Example
Digital art | Image makers | Clean up skies and backdrops from SFT renders | Render a foggy scene without stripes
Design | Visual designers | Keep smooth studio backgrounds even | Generate a plain wall backdrop
Workflow building | ComfyUI builders | Add one LoRA node to an existing graph | Insert the loader before sampling
Tooling | Advanced ComfyUI users | Use split checkpoints with bypass LoRA loading | Load a LoRA in bypass mode

If you only use the Base model, you can skip the LoRA, since it is built from SFT's own changes.

Where to use it: Everything here is meant for ComfyUI, which runs on Windows, macOS and Linux with Python.

• ComfyUI, with the LoRA in its loras folder
• ComfyUI, with the split checkpoints in its diffusion_models folder
• The LoraLoaderModelOnly node for normal use
• The Load LoRA (Bypass) node, with the split checkpoints

Use case | Who | What they get back
Remove banding | SFT users | Smooth areas without evenly spaced color stripes
Bypass loading | Advanced users | A working LoRA, since fused files drop q, k and v adapters
Base or SFT checkpoint | Workflow builders | Split q, k and v weights in bf16

How to set it up: You need ComfyUI, the MOSS-Image SFT model and Python with pip to fetch the files. The text encoder and VAE come from Comfy-Org's MOSS-Image repository and are unchanged.

• ComfyUI installed
• The SFT model, either from Comfy-Org or the split file here
• Room for 10.2 GB per split checkpoint

1. Install the Hugging Face hub package.
2. Download this repository.
3. Copy the LoRA into the loras folder and the checkpoints into diffusion_models.
4. Add the text encoder and VAE from Comfy-Org's repository.

pip install -U huggingface_hub
hf download nynxz/MOSS-Image

How to use it: Load the SFT model, then add the LoraLoaderModelOnly node with the destripe file. Set the strength by image width, then sample as usual. Nothing else in the workflow changes.

• 1.0 at 768 px wide
• 1.5 at 1024 px wide

Tool | What it does
LoraLoaderModelOnly | Applies the destripe LoRA to the SFT model
Load LoRA (Bypass) | Needs the split checkpoints from this repository

Example: You render a sky at 1024 px with the SFT model and it shows green and purple vertical stripes. You add the LoRA at strength 1.5 and render again with the same prompt and seed, and the sky comes back clean.

For newcomers and experts: If you are new, start with one prompt that has a big flat area, like a sky, render it with SFT, then add the LoRA and compare. The two images show what the file does.

If you know the field, the LoRA removes only the few directions inside the stripe-carrying layers, where another fix reverts whole layers. The card reports it keeps closer to SFT than that fix and is 1.9 MB.

Load the MOSS-Image SFT model
Add the LoRA loader node
Set strength by image width
Sample as usual

Good to know: Strength 1.0 at 1024 px can leave a slight trace in roughly half the images. Going up to 2.0 brings no gain over 1.5 and drifts further from the SFT look.

• License: apache-2.0
• Cost to run: none named
• Data it reaches: files you download, with no account or web service named
• Not for: the Base model, since the LoRA is built from SFT's changes

On Hugging Face, 13 November 2026

• Library: diffusion-single-file
• Downloads in the last 30 days: 6,362
• Likes: 21
• License: apache-2.0
• Based on: OpenMOSS-Team/MOSS-Image-Base
• Tags: diffusion-single-file, comfyui, lora

Links

• Model: https://huggingface.co/nynxz/MOSS-Image
• Files: https://huggingface.co/nynxz/MOSS-Image/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download nynxz/MOSS-Image

Then follow the usage in its model card. Nothing in this kit runs by itself.