MOSS-Image
#15 v Text na obrázekMoss · od 16. Oktober 2026 · 2× · naposledy 20. 10. 2026
19
Momentum
MOSS-Image is a text-to-image model from OpenMOSS with a 5B-parameter diffusion transformer (single-stream DiT, pretrained from scratch), a Qwen3.5-9B text encoder and a FLUX.2 VAE. It supports prompts up to 4K tokens and experimental JSON layout control with bounding boxes. Base and SFT checkpoints are released under Apache-2.0; Base is intended for fine-tuning/LoRA.
Vývoj momenta
23.07.21.10.
Vlastnosti
| Fine-Tuning | Base checkpoint for fine-tuning and LoRA; LoRA training also available in Civitai Training Studio |
| License | Apache-2.0 (weights and inference code) |
| Max Resolution | Recommended 1M, 1.5M, 2M pixels; extrapolates from 0.25M (512×512) to 4M (2K) |
| Platform | Open weights (Hugging Face, ModelScope); Python/Diffusers, Gradio, ComfyUI (core support) |
| Release Date | October 16, 2026 |