MOSS ImageNew

MOSS Image is a 5B-parameter diffusion transformer from OpenMOSS, the open-source lab at Fudan University. It pairs a Qwen3.5 9B text encoder with the FLUX.2 VAE, which is where its two headline abilities come from: the large language encoder reads prompts of up to 4K tokens in natural language, booru-style tags or a JSON layout, and the model was trained to place text and elements where the prompt says. It renders from 512px up to 2K at any aspect ratio and is text-to-image only; image editing and reference images are not part of this release.

Two versions are published. SFT is the supervised fine-tune of Base on curated data and the one to generate with: OpenMOSS’s own comparison shows cleaner text grouping and more organised layouts. Base is the undistilled base model with no post-training, the one to fine-tune and train LoRAs on. On Civitai the two share one ecosystem, so a LoRA trained on either loads on both. OpenMOSS recommends 50 steps with CFG 3 (raise it for short prompts), flow-matching Euler and shift 5; Civitai’s generator starts at 25 steps and goes up to 60. The weights and inference code are Apache-2.0: commercial use, derivatives and LoRAs are allowed, with the licence notice kept on redistribution.

Choose MOSS Image when the picture has to carry words, such as a poster, a label, a sign or a comic panel, or when a long, specific caption is easier to write than a tag soup. For anime with a deep tag vocabulary, Anima and Illustrious are the better-trodden paths; for photoreal editing, Krea 2 or FLUX.2. On Civitai both MOSS versions sit in the generator with steps, CFG and a negative prompt exposed, and a growing LoRA library trained on Base.