Fudan's MOSS-Image goes open source: 5B DiT with CN/EN text rendering, plus 8-step DMD LoRA
ostrisai · x · 2026-10-18
OpenMOSS has open-sourced MOSS-Image on Hugging Face, a 5B-parameter Diffusion Transformer (DiT) image model focused on Chinese & English text rendering and controllable layouts.
Key features:
- Long prompts & text rendering: up to 4K-token prompts including multi-line text, targeting posters and infographics
- Experimental JSON layout control: specify text/objects and positions via bounding boxes
- Flexible generation: multiple aspect ratios, resolutions, custom inference settings
- Developer workflows: Diffusers, Gradio, ComfyUI, plus a Prompt Enhancer Skill
The community also released an 8-step data-space DMD distillation LoRA (from the 50-step CFG 3 base down to 8 evals, no CFG). Load the LoRA on the stock transformer and sample with shift 5, CFG 1. Caveat: distillation weights must be loaded and run in fp32 — casting to bf16 destroys most of the distillation. Base and SFT checkpoints are both available under Apache 2.0.
More from Multimodal
- Reverse-engineered prompt schema unlocks reliable character swap in MiniMax H3 — photobombolo · 2026-10-18
- Black Forest Labs returns with new image model boasting strong editing, live on vLLM — Xianbao_QIAN · 2026-10-18
- DeepSeek V4.1 Flash one-shots a detailed 1990s Chinese apartment block in Blender — zizhpan · 2026-10-18
- Concat: open-source Rust-based CapCut alternative with a JSON-RPC API for AI agents — techNmak · 2026-10-18
- kimo: An Open-Source GUI-less Video Editor That AI Agents Drive via MCP — LieTop7965 · 2026-10-18
- Paying $0.50 per 10 seconds for video gen APIs — when is self-hosting worth it? — Key-Lychee-913 · 2026-10-18