Open-source image generation has a new breakout hit. MOSS-Image, a text-to-image model from MOSI and OpenMOSS, racked up roughly 1.3 million downloads in its first four days on Hugging Face, climbing to the top of the platform’s text-to-image trending board and picking up around 1,400 GitHub stars along the way.
What sets the model apart is control. MOSS-Image is a 5-billion-parameter diffusion transformer trained from scratch, and it is built for design work rather than generic pretty pictures. It accepts prompts of up to 4,096 tokens and takes structured JSON layout input with bounding boxes, so a headline, a product shot, and a background land exactly where the creator placed them. It also renders readable text in English and Chinese, a task where many image models still stumble.
The weights and code are released under the Apache 2.0 license, which means anyone can use them commercially. Community builds are already spreading fast: an 8-step distilled variant has passed 75,000 downloads, alongside GGUF, INT8, and MLX versions for different hardware. The model runs inside ComfyUI, the node-based workflow tool popular with power users.
The milestone underlines how quickly open image models are closing the gap on closed ones. For creators comparing options, Flux and Stable Diffusion remain the established open-weight choices, while Midjourney leads on the closed, hosted side. MOSS-Image’s edge is its layout precision, which matters most for posters, ads, and title cards where text placement is the whole job.
Source: RobotToday, SiliconNoon
