OpenMOSS releases MOSS-Image: 5B open-weight text-to-image model with JSON layout control
lmoroney · x · 2026-10-18
OpenMOSS released MOSS-Image on October 16, an open-weight (Apache 2.0) text-to-image model family.
- 5B backbone: a Base model plus an SFT version fine-tuned on curated data, built for efficient generation and fine-tuning; training used heterogeneous GPU platforms
- Native layout control: beyond natural-language prompts (up to 4K tokens) and tag sequences, a structured JSON prompt specifies each object or rendered text with a caption and bounding box, and the model composes the image to match
- Strong text rendering, covering both anime and photorealistic styles
- Ecosystem: Diffusers custom pipeline (PyTorch 2.9+), a Gradio app, a ComfyUI plugin in ComfyUI Manager, with official integration in progress
Related event: Fudan's OpenMOSS Open-Sources 5B MOSS-Image Text-to-Image Models(14 posts)→
More from Multimodal
- 'The Glitch': A Robot Folk Horror Short With a Cracked-Doll Android — BotLordBotflix · 2026-10-19
- tscomfy: Type-Safe ComfyUI Workflows as Code, Types Generated From Your Server — No_One_9304 · 2026-10-19
- 'Connect' Short Film Shows AI Video Closing In on Photoreal Human Eyes — indiegameplus · 2026-10-19
- MOSS image model design test: consistent pig character cards across six actions — DevKkw · 2026-10-19
- French #slopcore AI track went viral without lyrics, now gets an English-subtitled version — banteg · 2026-10-19
- AI-Generated Claymation-Style Horror Short 'Satanic Verse' Turns Heads on Reddit — Pnarpok · 2026-10-19