Summary
Moss Image is a new 5-billion-parameter image generation model from Open Moss, the team behind Moss TTS, and it is released under the permissive Apache 2.0 license. Sam Witteveen walks through what it can do, where it falls short, and why a small open model is drawing attention.
The model is a diffusion transformer trained from scratch and paired with a Qwen 3.5 9B text encoder, which allows prompts of around 4,000 tokens. Two checkpoints are available: a base model for fine-tuning and LoRA training, and an SFT model for generation. The video shows strong photorealistic portraits, product shots, lighting, and varied styles, with resolutions from about 0.25 to 4 megapixels. Text rendering is the standout feature, covering multi-line posters, chalkboard menus, infographics, and English and Chinese in the same scene. The developers’ own long-text benchmark places it just behind Nano Banana and GPT Image 2, though no third-party evaluation exists yet.
A JSON-based layout feature lets users specify where text goes in the image, and an LLM can generate that JSON from a plain description. Limitations include no image-to-image editing, slow default settings of 50 steps, and breakdowns when too much text is requested. The community has already released 4-step and 8-step LoRAs, quantized versions, and a fine-tune, and an official 4-step distillation is in development.
📺 Source: Sam Witteveen · Published October 26, 2026
🏷️ Format: Review
