arXiv 2610.23374
MOSS-Image Technical Report
AI13 upvotes
Authors: MOSI. AI
Overview
The authors present MOSS-Image, an open generation system that pairs a 5B single-stream diffusion transformer (DiT), pretrained from scratch, with an intention-aware agentic prompt enhancement module. Together they target bilingual long-form text rendering and precise layout control. They also introduce two benchmarks, LongLongTextBench (LLTB) and RealLayoutArena (RLA).
Introduction
Closed-source image generators produce diverse content, but open-weight text-to-image models still struggle with dense text and intricate layouts.
Results
MOSS-Image achieves state-of-the-art performance among open-weight models on LLTB, with an average text coverage of 96.3%. On RLA it ranks second in end-to-end system comparison, and its backbone ranks first among generators evaluated with shared enhanced prompts.
Conclusion
The planned release of the model, training recipe, and evaluation framework gives the open community a stronger base for text-rich, layout-aware image generation.