An independent developer publishing as chinoll on Hugging Face and Chenkin-x on GitHub has released MOSS Image Edit, a LoRA adapter that introduces instruction-based, multi-reference image editing to OpenMOSS's 5-billion-parameter image models. The underlying MOSS-Image model, detailed in a technical report published by MOSI.AI on arXiv on October 16, 2026, is a 5-billion-parameter single-stream diffusion transformer trained from scratch. Previously, the official OpenMOSS models supported text-to-image generation and structured layout editing via bounding boxes, but lacked direct plain-language editing capabilities such as changing seasons, lighting, or outfits from reference photos.

Technical Architecture

According to the project documentation, the adapter introduces a reference branch extending the MOSS backbone where tokens are ordered as text, reference, and target. Reference images remain noise-free at a fixed timestep of zero. The reference branch runs once, caching layer keys and values to reuse during every denoising step, which helps optimize computation. The adapter features rank-64, alpha-64 weights in bf16, applied across all 42 blocks of the network. It operates within ComfyUI, utilizing default settings that output 512x512 images using 30 steps at a classifier-free guidance of 4.0, supporting up to four reference images.

Licensing and Community Reception

The licensing structure remains ambiguous. While the project code is distributed under the GPLv3 license, neither the model card nor the repository specifies a dedicated license for the LoRA weights themselves, while the base models retain an Apache-2.0 license. Furthermore, no formal quantitative benchmarks have been published for the editing capabilities. Independent testing by community developers indicates mixed results, particularly when applying the LoRA to the base model instead of the SFT version. Currently, the release remains a hobbyist-level contribution while the community awaits official ComfyUI integration from OpenMOSS.

Source: AI Daily News. This story was rewritten and translated for our readers.