MOSS Image Edit: Run the Community LoRA in ComfyUI

MOSS-Image has no official edit model — but a community LoRA adds text-instruction editing, and ComfyUI merged native support for it on October 22.

This page covers both routes, and one thing up front: this site cannot upload your photo. Our generator accepts text and layout only.

What you can do here is edit the text and layout of an image you generated on this site, then re-render it with the same seed.

What it can doWhat goes inInstall needed?
Official MOSS-Image Base / SFTText-to-image, including JSON layouts that place text and objectsA text prompt or a JSON layoutYes, to run it yourself (ComfyUI, Diffusers or the Gradio demo)
Community edit LoRA (chinoll/Moss-Image-Edit)Changes an existing image from a text instruction: colors, outfits, lighting, seasons, visual styleOne to four reference images and an instructionYes: ComfyUI, the LoRA and a MOSS-Image Base or SFT model
Layout editing on this siteRe-renders an image you generated here after you edit its text or layout, with the same seedAn image generated on this site, with its stored layout and seedNo, it runs on this site

What the community LoRA actually does

chinoll/Moss-Image-Edit is an image-editing LoRA for MOSS-Image, published on Hugging Face on October 22, 2026. It is a community project, not an OpenMOSS release. Its model card says:

  • What it changes: colors, outfits, lighting, seasons and visual style, described in a text instruction.
  • What it starts from: an image you already have. Up to four images can be connected; with more than one, the instruction should say what each image is for.
  • Which base model: MOSS-Image-Base or MOSS-Image-SFT, either one. The author's ComfyUI guide adds that the patch was trained against SFT and suggests starting with SFT; Base loads too, but you have to judge its editing quality yourself. Our Base vs SFT guide explains how the two checkpoints differ.
An illustration of a girl reaching toward a bright blue sky, labeled Original, and the same scene with a pink and orange sunset sky, labeled Edited
From the chinoll/Moss-Image-Edit model card. Its full showcase shows more edits, each next to its original.

The same model card is published on ModelScope as ChenkinLab/Moss-Image-Edit, and the training code is on GitHub at Chenkin-x/Moss-Image-Edit. A separate community repository, akhaliq/MOSS-Image-Edit-HQ, publishes several more edit LoRA versions; its card describes one file in the same format as chinoll's adapter and another in the ai-toolkit format.

Install the LoRA

The model card gives these download commands, using SFT as the example:

hf download chinoll/Moss-Image-Edit adapter_config.json adapter_model.safetensors --local-dir models/Moss-Image-Edit
hf download OpenMOSS-Team/MOSS-Image-SFT --local-dir models/MOSS-Image-SFT

For Base, it gives this base-model command instead:

hf download OpenMOSS-Team/MOSS-Image-Base --local-dir models/MOSS-Image-Base

For installation and model placement, the model card links to the author's ComfyUI guide. That guide puts the files here:

FileDirectoryNative loader
ComfyUI-format MOSS Image SFT / Basemodels/diffusion_models/Load Diffusion Model
Qwen3.5-9B text encodermodels/text_encoders/Load CLIP, type moss_image
Matching FLUX.2 VAEmodels/vae/Load VAE
MOSS Edit adapter_model.safetensorsmodels/model_patches/Load Model Patch

The two pages differ in one place. The model card's commands fetch the full Hugging Face repositories, and it says to point a ComfyUI loader's snapshot_dir at the base model. The current guide uses ComfyUI's single-file models instead and says a full Diffusers snapshot is unnecessary. If you follow the guide, take the base model, text encoder and VAE in ComfyUI format: our MOSS-Image ComfyUI workflow guide lists those files and where they go. The edit patch is the adapter_model.safetensors file from the model card, and the guide notes you may rename it to moss_image_edit.safetensors.

Two routes inside ComfyUI

Native nodes: ComfyUI v0.43.0 or later

Kijai's pull request #17296, "Support MOSS-Image edit LoRAs", was merged on October 22, 2026. The first release that includes it is v0.43.0, published on October 27. It adds an experimental node, Text Encode MOSS-Image Edit:

  • Set edit_format to Moss-Image-Edit for this LoRA. The author's ComfyUI guide gives the same instruction.
  • Connect the VAE and your images. With this format the references reach the model only as reference latents, and the text encoder sees your instruction alone, so the instruction has to describe the change in words.
  • Use a second copy of the node, with the same images, for the negative conditioning.
  • Scale the images first. The node uses them at the size they arrive.

The LoRA file goes through ComfyUI's regular LoRA loader: the same pull request maps the key names that Moss-Image-Edit files use. akhaliq's model card describes the native graph as Load Diffusion Model → LoRA loader → VAE loader → Text Encode MOSS-Image Edit (positive and negative) → KSampler → VAE Decode, with the single-file weights from Comfy-Org/MOSS-Image.

The author's custom nodes

Chenkin-x/ComfyUI-Moss-Image-Edit adds an Apply MOSS Edit node. It takes the model, the edit patch, the VAE, the positive and negative conditioning, a source image and optional extra references. Per its guide:

  • It needs ComfyUI with native MOSS-Image support. The plugin's tested baseline is v0.42.3.
  • Install it as a custom node, then run install.cmd (Windows) or install.py with the Python that starts ComfyUI. That script only checks your environment; it installs nothing.
cd ComfyUI/custom_nodes
git clone https://github.com/Chenkin-x/ComfyUI-Moss-Image-Edit.git
  • Place the files as in the table above. The edit patch loads with Load Model Patch.
  • Start from its single-reference SFT workflow. Its starting settings are a 512-pixel limit on the reference's long edge, about 1 MP output, Euler / normal, 30 steps and CFG 4.

Not for this LoRA: ostris's nodes

ostris/ComfyUI-MOSS-Image-Ostris-Edit is for edit LoRAs trained with ai-toolkit's MOSS-Image edit training, not for chinoll's file. Its Text Encode MOSS-Image Ostris Edit node takes clip, prompt, an optional vae and up to three images, and it also needs ComfyUI's native MOSS-Image support.

One reference or four?

The model card allows up to four images and asks you to explain each image's role in the instruction. The author's ComfyUI guide is more careful: it says multi-image editing is implemented but that the current LoRA's quality with several images needs evaluation, and it recommends SFT single-image editing as the starting point.

So start with one image. Once a single-image edit works the way you want, add a second reference and say in the instruction what it is for.

What you can do here without installing anything

This site does not run the edit LoRA, and it cannot take your photo. What it can do is let you edit the layout of an image you generated here:

  1. Generate an image with the MOSS Image AI Generator.
  2. If the image has a stored layout, an Edit layout button appears under it. It opens the layout editor with your image underneath its boxes and its seed kept.
  3. Change a box's text or description, or move it, then click Regenerate to re-render with the same seed, or New look to try a different one.

This is a re-render, not a pixel edit. With the layout and seed unchanged, it reproduces the image. Once you change something, the whole image is drawn again, and the parts you did not touch may not stay exactly the same.

Known gotchas

  • No Text Encode MOSS-Image Edit node. Your ComfyUI is older than the change. The node is in v0.43.0 and later; the v0.42 releases do not have it.
  • The model card and the guide load the LoRA differently. The card still mentions a loader's snapshot_dir; the author's current ComfyUI guide uses ComfyUI's single-file models and Load Model Patch. Follow the guide.
  • The instruction is the only text the encoder sees. With edit_format set to Moss-Image-Edit, your images reach the model as latents only, so describe the change in words.
  • The license of the weights is not stated. The model card does not name one. The training code is under GPLv3. Check the terms yourself before using results commercially.
  • Quality with several references needs evaluation, in the author's words. Start with one image.

Editing your own photo — not supported yet

This site cannot take an uploaded photo, and there is no date for when it will. If you want one email if it ever can, leave your address below. It is used for that email and nothing else.

One email if it happens, nothing else. See the privacy policy.

FAQ

Can I upload my own photo?

No. The generator on this site accepts only text prompts and layouts, so there is no way to upload an image; the only file you can import is a layout JSON in the layout editor. To edit a photo with MOSS-Image today, run the community edit LoRA in ComfyUI.

Is there an official MOSS-Image edit model?

No. OpenMOSS has released two checkpoints, MOSS-Image-Base and MOSS-Image-SFT. The edit LoRA chinoll/Moss-Image-Edit on Hugging Face is a community project.

Which ComfyUI version has the native edit node?

Native support was merged on October 22, 2026, in pull request #17296, and the first release that includes it is v0.43.0. Older releases have no Text Encode MOSS-Image Edit node. The author's custom nodes were tested from v0.42.3.

Does the edit LoRA work with Base or SFT?

The model card says either one works. The author's ComfyUI guide adds that the patch was trained against SFT, so start with SFT and judge Base yourself.

Can I use the edit LoRA commercially?

The model card does not state a license for the weights, and the training code is under GPLv3. Check the terms yourself, or ask the author, before using results commercially.