Moss-Image-Edit LoRA for ComfyUI

Moss-Image-Edit LoRA for ComfyUI

Edits an image, or combines up to four, from a text instruction, using a ComfyUI-ready version of the Moss-Image-Edit LoRA. For ComfyUI users working with the MOSS-Image model.

Made by PedroMarinhoDev 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 is an image-editing LoRA for the MOSS-Image model, with its key names changed so ComfyUI's built-in MOSS-Image support can load it. You give it one image, or up to four to combine, plus a text instruction, and it edits the picture. The original LoRA comes from chinoll and was trained on MOSS-Image SFT.

Who it suits: It suits ComfyUI users who want to change a picture by describing the change in words. Some familiarity with ComfyUI folders and nodes helps.

Industry | Who uses it | How they use it | Example
Design | Designers who work in ComfyUI | Edit a photo by typing an instruction | Removing one object from a product photo
Content | Creators making social images | Change scene elements without redrawing the picture | Swapping a background element
Imaging | People who combine pictures | Merge two or more source images in one edit | Blending two reference photos
Research | Model testers | Compare this LoRA with the HQ versions the node pack also runs | Running one prompt on each version

You can skip it if you do not use ComfyUI or you edit images some other way.

Where to use it:

• ComfyUI 0.42 releases and ComfyUI master, through the ComfyUI-Moss-Image-Edit-Native node pack
• ComfyUI master on its own, which can run this LoRA but is not in a release yet
• A GPU, since the settings were tested on an RTX 3090
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
Edit one image | Designers | The same picture changed as the instruction says
Combine images | Creators | One result built from up to four sources
Compare LoRAs | Testers | Side-by-side results from this LoRA and the HQ ones

How to set it up:

• A working ComfyUI, either a 0.42 release or master
• The ComfyUI-Moss-Image-Edit-Native node pack
• Python and pip, for the download tool
• The base files listed below, which come from Comfy-Org/MOSS-Image

1. Install the node pack. It uses stock ComfyUI loaders, so it needs no diffusers or transformers.
2. Download the LoRA file into your ComfyUI loras folder with the command below.
3. Put the diffusion model, text encoder and VAE files in their folders.
4. Open a workflow from Templates, under the node pack's name.

Folder | File
loras | moss_image_edit_lora_bf16.safetensors, from this repository
diffusion_models | The MOSS-Image SFT model, in its bf16 or int8 form
text_encoders | qwen3.5_9b_w4a8_gptq, or a bf16 Qwen3.5-9B
vae | flux2-vae_small, or flux2-vae

pip install -U huggingface_hub
hf download PedroMarinhoDev/Moss-Image-Edit-ComfyUI moss_image_edit_lora_bf16.safetensors --local-dir ComfyUI/models/loras

How to use it: Open the template workflow, load your image or images, type your instruction and run it. The card's tested settings matter, since the LoRA was trained under narrow conditions.

Setting | What to use
LoRA strength | 1.0, because lower values redraw the scene instead of editing it
Reference and output size | The same size, which the node defaults handle for you
Sampler | CFG 5 over 30 steps with euler and simple. In the negative pass, leave out layer 22 and keep the prompt empty
Base model | MOSS-Image SFT, not Base
kv_cache | Auto, the default, which computes the reference images once per run

Example: Load a photo of a table with a lamp on it and ask for the lamp to be removed, with LoRA strength at 1.0. The card lists removals among the tested edit types. You should get the picture back with the lamp gone.

For newcomers and experts: A LoRA is a small add-on file that teaches a larger image model one extra skill. Here the skill is editing a picture from a typed instruction. For a first try, follow the setup, open the template and make one simple edit on one image.

Underneath, only the key names changed, and the tensors keep the original bf16 values. The file uses rank 64 with alpha 64 per module, so the scale is 1.0, and it targets the attention and feed-forward linears of all 42 blocks. The LoRA was trained with sources and targets at one size, which is why mismatched sizes zoom out or fail. Training used the SFT model, and the card reports that Base fails more often on text, swapping a subject and two-image edits.

Source images and a text instruction go in
The MOSS-Image SFT model loads the LoRA at 1.0
The sampler runs 30 steps at CFG 5
An edited image comes out

Good to know: Changing a pose and swapping one subject for another work poorly, and a swap tends to add the new subject next to the old one. With CFG set to 4, a removal can take neighbouring objects too.

• License: GPL-3.0, and the original repository names no separate license for the weights
• Cost to run: no API key or paid service is named, and you need a GPU that can run MOSS-Image
• Data it reaches: downloads model files from Hugging Face, and it edits the images you load in ComfyUI
• Not for: pose changes, subject replacement, or use on the Base model, where the card reports more failures

On Hugging Face, 14 November 2026

• Task: image to image
• Library: comfyui
• Downloads in the last 30 days: 0
• Likes: 18
• License: gpl-3.0
• Based on: OpenMOSS-Team/MOSS-Image-SFT
• Tags: comfyui, lora, image-editing, moss-image, image-to-image

Links

• Model: https://huggingface.co/PedroMarinhoDev/Moss-Image-Edit-ComfyUI
• Files: https://huggingface.co/PedroMarinhoDev/Moss-Image-Edit-ComfyUI/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download PedroMarinhoDev/Moss-Image-Edit-ComfyUI

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