OpenMOSS Releases MOSS-Image for Long Text and Structured Layouts
OpenMOSS, a team involving Fudan NLP and Shanghai Chuangzhi, released the open-source MOSS-Image image-generation model. Its 5B-parameter single-stream DiT backbone reportedly supports prompts of up to about 4K tokens, Chinese and English long-text rendering, and editable JSON layouts. Base and SFT weights are available under Apache-2.0, with integrations for Diffusers, Gradio, and ComfyUI. The model uses an external prompt-enhancement planner and cannot directly accept reference images or perform native image editing.
MOSS-Image provides open weights and tools for developers working on typography- and layout-sensitive posters, infographics, comics, and UI designs, including workflows requiring local deployment and reproducible layouts.
近日，OpenMOSS（复旦NLP与上海创智等团队）正式发布了开源图像生成模型 —— MOSS-Image。针对传统文生图模型在海报、信息图、漫画及UI设计中常年面临的“文字拼写错误、段落截断、元素错位”等痛点，该模型以5B参数的单流 DiT(Diffusion Transformer)为基座从零开始预训练，全面支持最长约4K token的提示词、中英文长文本渲染以及可编辑的结构化 JSON Layout。
在技术架构方面，MOSS-Image采用Muon–AdamW训练配方，通过统一的文本条件通道来处理自然语言和结构化布局。模型使用 Qwen3.5-9B 作为文本编码器，并配合 FLUX.2VAE 进行图像解码。训练数据集涵盖约3B张图像，融入了多粒度caption、文字监督与结构化边界框（bbox）监督。为了应对复杂创作需求，模型还引入了意图感知的提示增强(PE)外部规划模块，可将简短的请求自动扩展为包含字体层级、色板和坐标规范的可执行指令。
#MOSS-Image#OpenMOSS#Fudan NLP#Shanghai Chuangzhi#Qwen3.5-9B#FLUX.2VAE#Diffusers#Gradio#ComfyUI