MOSS Image Edit LoRA adds instruction editing to OpenMOSS model

A developer publishing as chinoll on Hugging Face and Chenkin-x on GitHub has released MOSS Image Edit, a LoRA adapter that lets OpenMOSS's 5-billion-parameter image models edit pictures from plain-language instructions. The model card describes it as "an image-editing LoRA for MOSS-Image" that changes "colors, outfits, lighting, seasons, and visual style with text instructions." It works with both MOSS-Image-Base and MOSS-Image-SFT. Edits run in ComfyUI from a single reference image or up to four.

The release fills a visible gap. OpenMOSS shipped MOSS-Image-SFT for text-to-image and JSON layout generation. Its only editing route was structured: users rewrite text, element descriptions or bounding boxes in a layout file. Telling the model to "make it winter" or to put a person into a new outfit from a photo was not part of the official offering. Now a small community adapter supplies it.

A cached reference branch on a 5B DiT

MOSS-Image itself is recent. The MOSS-Image technical report, credited to MOSI.AI and posted to arXiv on October 16, 2026, describes a 5B single-stream diffusion transformer trained from scratch. On LongLongTextBench, a benchmark the authors built themselves, it averaged 96.3% text coverage, which they call the best result among open-weight models. OpenMOSS's GitHub repo pitches the model for text rendering, layout generation and "efficient fine-tuning." Developers appear to have taken up that pitch. The SFT model tree listed roughly 30 adapters when AI Daily checked, a live figure that shifted between page loads.

Chenkin-x's design is straightforward. According to the project README, a reference branch extends the MOSS backbone, with tokens ordered as text, reference, then target. Reference images stay noise-free at a fixed timestep of zero. The branch runs once, stores each layer's keys and values, and reuses them at every denoising step. "Since the reference branch does not depend on text, the same cache also serves both positive and negative CFG branches, reducing repeated computation," the README says.

The adapter itself is modest. A community conversion by developer PedroMarinhoDev reports rank-64, alpha-64 weights in bf16. They apply to the attention and feed-forward projections across all 42 blocks, or 252 linear layers. The README also lists facial expressions and scenes as editable, and it promises users can "combine people, outfits, and scenes using up to 4 reference images."

Shipping defaults are conservative. The ComfyUI node pack outputs 512×512 images, caps reference edges at 512 pixels and runs 30 steps at CFG 4.0. Its multi-reference workflow wires up two images by default. Users must add the third and fourth inputs themselves.

The GPLv3 label covers code, not weights

The licensing is less clear than it first looks. The Hugging Face card says "project code is available under GPLv3," and the GitHub repository carries a GPL-3.0 badge. The README adds that "base models and third-party dependencies retain their respective licenses." Nothing in either repo states a license for the LoRA weights themselves, and the chinoll Hugging Face page has no license tag. PedroMarinhoDev flagged the gap directly: "The original model repo names no separate license for the weights; this copy follows the project's GPL-3.0."

That gap matters for anyone weighing commercial use. The base models are Apache-2.0. GPLv3 has copyleft terms, and applying them to a set of tensors is legally untested. An unlicensed weight file leaves users guessing.

Ghosts on Base, and no benchmarks

Neither the model card nor the README publishes a quantitative editing benchmark. The only evidence is the showcase images, credited to GitHub user @onglon114514. The only outside testing on record comes from PedroMarinhoDev, and his results are mixed. He found the LoRA "was trained on SFT and also loads on Base." Multi-image work suffers on Base, though: "every seed tried left a ghost of image 2 where it sits in its own frame," he wrote, and the cleanest result came from the seventh seed. He also reports that at the node pack's default strength of 0.25, two-image edits return only the second image. Strength 1.0 is needed.

Traction so far is thin. The chinoll repo had a single like when checked, and the GitHub project had five stars. This is a hobbyist-scale release, not a lab product.

The next signal to watch is whether OpenMOSS's native ComfyUI integration, which its SFT card says "is in progress," absorbs community converters like PedroMarinhoDev's. Also watch whether chinoll clarifies the weight license or publishes editing benchmarks. Until then, MOSS Image Edit is a promising adapter with no published editing scores.