The OpenMOSS team has released the open-source image generation model MOSS-Image, built on a 5B-parameter single-stream DiT architecture. It supports prompt lengths of up to 4K tokens and high-quality rendering of long-form Chinese and English text, with editable structured JSON layout capabilities. The model excels in long-text rendering, layout accuracy, and end-to-end preference evaluations. Released under the Apache-2.0 license, it is compatible with the Diffusers, Gradio, and ComfyUI ecosystems, making it ideal for developers requiring precise typography and local deployment.Author and source: AIBase
Recently, OpenMOSS (a collaboration between Fudan NLP, Shanghai Chuangzhi, and other teams) officially released the open-source image generation model—MOSS-Image. Addressing long-standing pain points in traditional text-to-image models—such as misspelled text, truncated paragraphs, and misaligned elements in posters, infographics, comics, and UI design—this model is pretrained from scratch using a single-stream DiT (Diffusion Transformer) with 5 billion parameters, fully supporting prompts of up to approximately 4K tokens, rendering of long-form Chinese and English text, and editable structured JSON layouts.
In terms of technical architecture, MOSS-Image employs the Muon–AdamW training recipe, processing natural language and structured layouts through a unified text-conditioning channel. The model uses Qwen3.5-9B as the text encoder and pairs it with FLUX.2VAE for image decoding. The training dataset encompasses approximately 3 billion images, incorporating multi-granularity captions, text supervision, and structured bounding box (bbox) supervision. To address complex creative demands, the model integrates an intention-aware Prompt Enhancement (PE) external planning module that automatically expands brief requests into executable instructions specifying font hierarchies, color palettes, and coordinate specifications.
In terms of core capabilities and evaluation performance, MOSS-Image demonstrates strong specialized advantages:
Long-text rendering (LLTB): In an evaluation containing 120 tasks each in Chinese and English, with an average target text length of approximately 1,339.8 characters, it achieved a Mean-6 score of 96.3, setting a new record for open-source weighted models and effectively preserving line breaks and paragraph structure.
Structured layout (7Bench): On the unified JSON layout task, it achieves an mIoU of 0.8786 for bounding boxes and an mIoU of 0.6788 for overall layout, demonstrating excellent performance in spatial relationships and relative size control.
End-to-end preference (RealLayoutArena): Its generator achieved a score of 1807.34 in blind Elo rankings under unified enhanced prompts, placing it at the forefront of the industry.
Currently, the Base and SFT weights of MOSS-Image are open-sourced under the Apache-2.0 license and fully integrated with ecosystem tools such as Diffusers, Gradio, and ComfyUI, supporting both LoRA and full fine-tuning. However, the official documentation clearly defines its current limitations: its backbone is a pure text-to-image model that cannot directly accept reference images or perform native image editing, and its prompt enhancement module relies on external language models. Overall, the release of this model provides developers requiring local deployment, precise layout, and reproducible formatting with a powerful open-source option.