Introduction

MOSS-Image can now edit images, not just generate them. MOSS Image Edit is an open-weights image-editing LoRA from community developer Chenkin-x (published on Hugging Face as chinoll/Moss-Image-Edit on October 22). It sits on top of OpenMOSS's 5B MOSS-Image Base or SFT checkpoint and lets you change colors, outfits, lighting, seasons and style with a plain-language instruction, using one to four reference images.

You can run it today in ComfyUI two ways: with the author's custom nodes and workflow JSONs, or with the official nodes Comfy merged into core on October 22 (PR #17296). It earned about 40 likes on Hugging Face in its first day, which is a lot for a single LoRA.

What Shipped

  • Weights: adapter_model.safetensors plus its config on Hugging Face. It works with either MOSS-Image-Base or MOSS-Image-SFT.
  • Edits: text-instruction changes to expressions, clothing, colors, objects, lighting, seasons and visual style, with the rest of the scene kept.
  • References: one source image, or up to four references whose roles you describe in the prompt (for example, the person from picture 1 in the jacket from picture 2).
  • Training code: open on GitHub under GPL-3.0, so you can train your own edit LoRA. The Hugging Face card doesn't list a license for the weights, so check before you use them commercially.

How It Works

The clever bit is how the references are handled. Reference images go through the same MOSS transformer as the image being generated, but they're kept noise-free at a fixed timestep of 0, and they only attend to each other. Because the reference branch never changes, it runs once, and its keys and values are cached and reused on every denoising step. That means extra references don't get re-processed on every step, which matters when you're stacking four of them.

Running It in ComfyUI

Native route (newest ComfyUI from master): PR #17296 teaches core ComfyUI to feed reference latents into MOSS-Image. It adds an experimental Text Encode MOSS-Image Edit node with an edit_format switch that covers this LoRA (set it to Moss-Image-Edit) and edit LoRAs trained with ai-toolkit. As of 17:00 MYT the change is on master but not yet in the tagged v0.42.5 release, so use a nightly or update from git.

Custom-node route: ComfyUI-Moss-Image-Edit has an Apply MOSS Edit node that works with the stock loaders and KSampler, plus three example workflows: SFT single image (the recommended start), Base BF16 and multi-reference. The author notes that multi-reference quality "needs evaluation," so start with single-image edits.

Grab the files (SFT shown):

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

Try It Yourself

Name what changes, then name what stays. A good single-image starter:

Change the season to deep winter with soft snowfall and frosted windows, keep the person's face, pose and outfit exactly the same.

And one for two references:

Put the woman from Picture 1 into the denim jacket from Picture 2, keep her hairstyle, expression and the original café background unchanged.

Conclusion

MOSS-Image has had a busy week: native Comfy support, fast LoRAs, color and depth control, and now instruction editing. An edit LoRA that runs inside a stock ComfyUI graph makes this open 5B model a lot more useful day to day. Expect more edit LoRAs now that core has an edit_format switch waiting for them, and keep an eye out for a tagged ComfyUI release that includes PR #17296.

—Aurelia ♡