MOSS-Image
MOSI.AI published the MOSS-Image technical report, describing a 5B single-stream diffusion transformer for text-rich, layout-heavy images with an agentic prompt enhancement module.
MOSS-Image pretrains a 5B single-stream DiT from scratch for bilingual long-form text rendering and layout control. It adds an intention-aware agentic prompt enhancement module that turns user requests into detailed content and layout specifications. The authors also introduce two benchmarks, LongLongTextBench for long English and Chinese text rendering and RealLayoutArena for layout-rich generation.
- Date
- Friday 16 October 2026
- Lab
- MOSI.AI
- Kind
- paper
- Access
- paper only
Figures
| Measure | Value | Measured by |
|---|---|---|
| LongLongTextBench average text coverage | 96.3% best among open-weight models, per the authors | company |
| RealLayoutArena end-to-end system rank | 2nd its backbone ranks 1st among generators using shared enhanced prompts | company |
| Model size | 5B parameters single-stream DiT trained from scratch | company |
The paper says the model, training recipe and evaluation framework will be released, so weights were not confirmed available at the time of the report.
Sources
This record was checked and corrected against its sources on 19 October 2026. How we check