How do I edit images with MOSS-Image Edit?
Last updated 2026-10-27
MOSS-Image Edit is a LoRA that adds instruction-based image editing to MOSS-Image. ComfyUI added native support on Oct 22, 2026 (PR #17296): update past v0.42.1, load a MOSS-Image checkpoint, add an edit LoRA and use the Text Encode MOSS-Image Edit node. 50 steps, CFG 3, Euler, skip-layer 22. Up to 4 reference images.
Which MOSS-Image Edit LoRA should I use?
| LoRA | Author / release | Format | Best for |
|---|---|---|---|
| moss_image_edit_lora_rank_64_bf16 | Packaged by Comfy-Org; 0.3 GB | Native ComfyUI template | Default choice |
| chinoll/Moss-Image-Edit | Community, 2026-10-22 | edit_format=Moss-Image-Edit; single-image and multi-image (up to 4) workflow JSON | Color, outfit, lighting, season, style |
| akhaliq/MOSS-Image-Edit-HQ (armV5) | Community, 2026-10-22 | edit_format=ai-toolkit; an 8-step Fast mode Space exists | Fidelity. armV6 has color casts; not recommended |
How do I run MOSS-Image Edit in ComfyUI?
- Update ComfyUI to a version past v0.42.1.
- Open Workflow → Browse Templates → MOSS-Image Edit.
- Put the edit LoRA file in
models/loras. - Upload the reference image and write the instruction.
- Run with 50 steps and CFG 3.
File names, sizes and VRAM options are on the ComfyUI workflow page. The ComfyUI docs cover the node setup.
How do I write an edit instruction?
- One sentence, one change.
- With several images, say what each one does, for example: "use image 2's jacket on the person in image 1".
- ComfyUI shrinks reference images automatically to 384×384 or smaller.
| What to change | Example instruction | |
|---|---|---|
| Background swap | Replace the background with a clean white studio backdrop and soft shadow. | Try → |
| Background swap | Move the subject to a sunny beach and keep the lighting on the subject natural. | Try → |
| Color change | Change the jacket to dark green leather; keep the face and pose unchanged. | Try → |
| Outfit change | Change the shirt to a white linen shirt and keep everything else the same. | Try → |
| Day to night | Make it night with warm street lights and a light snowfall. | Try → |
| Season change | Change the season to autumn with orange leaves on the trees. | Try → |
| Anime style | Turn this photo into a clean anime illustration with flat colors. | Try → |
| Multi-image reference | Use image 2's jacket on the person in image 1. | Try → |
How does MOSS-Image Edit compare with Fibo-2 Edit?
The akhaliq model card includes side-by-side images against Fibo 2.0 Edit. This page links to them and does not copy them.
This site's editor uses Bria Fibo-2 Edit (bria/fibo-2/edit on fal.ai). It takes an instruction plus reference images. It is not MOSS-Image Edit.
Where can I edit without a GPU?
Use the online editor: 3 free images, no signup. Background, color, lighting and style edits run on Bria Fibo-2. The Text poster mode runs on MOSS-Image. Prompt help is in the prompt guide.
Where do I get the models?
- OpenMOSS/MOSS-Image on GitHub: code and user guide.
- Comfy-Org/MOSS-Image on Hugging Face: ComfyUI files, including the edit LoRA.
- chinoll/Moss-Image-Edit and akhaliq/MOSS-Image-Edit-HQ: community LoRAs.
FAQ
Is MOSS-Image Edit an official OpenMOSS model?
No. The edit LoRAs on Hugging Face are community releases. Comfy-Org packages a rank-64 edit LoRA for ComfyUI's native template, and the OpenMOSS team has said it is exploring an official edit model.
Can I use MOSS-Image Edit online?
Not on this site. The online editor here runs edits on Bria Fibo-2 via fal.ai because fal.ai lists only the text-to-image MOSS-Image model. To run the MOSS-Image Edit LoRA you need ComfyUI on your own GPU, or a community Hugging Face Space.
How much VRAM does it need?
The sources list file sizes, not VRAM. The bf16 diffusion model and text encoder total about 29.5 GB on disk; the int8 diffusion model plus the w4a8 text encoder total about 12.8 GB. Use the smaller pair if you run out of memory.
Can I use the results commercially?
MOSS-Image is released under Apache-2.0. The chinoll repository code is listed as GPLv3, so check the license of each LoRA and workflow you use before commercial work.
Why do faces or colors shift in my edits?
Community reports on the akhaliq model page describe oversaturated colors, and the armV6 checkpoint has a color cast. Use armV5 with 50 steps.