MossYume
a new anime-and-realism base model for image generation
2 min read
MossYume is a text-to-image generation model fine-tuned from MossImage-base by OpenMOSS-Team. Version 0.1 was trained on a dataset of 20 million hand-curated images—with zero AI-generated or distilled data—maintaining a 1:1 balance between anime and realistic styles. It supports dynamic resolutions from 1024px up to native 2048x2048 high quality. Because this is a base release without SFT (supervised fine-tuning) or RL (reinforcement learning), detailed, complex prompts are recommended for the best results.
Key features
- Dataset: 20 million images from public sources (Danbooru, Pixiv, Gelbooru, e621, Konachan) with no synthetic data.
- Style balance: an equal split between anime and realistic images in the training set.
- Native 2K: full support for high-resolution generation up to 2048×2048 pixels.
- Architecture: designed to fix anatomical inconsistencies and stabilize the weights of the original MOSS model.
What you can do
- High-res generation: create detailed artwork and realistic shots at up to 2K resolution.
- Prompt experimentation: use complex, detailed text descriptions to unlock the base model's full potential.
- Pipeline assembly: hook up an external VAE and text encoder to run a complete generation workflow.
- Run in int8: use the quantized version to save VRAM.
Example
Use detailed English prompts to generate intricate anime or realistic images at resolutions up to 2048×2048.
Read more
Details• Dataset: 20M manually curated images (Danbooru, Pixiv, Gelbooru, e621, Konachan)
• Resolution: 1024px to 2048×2048px
• Components: diffusion model only (requires external VAE and text encoder)
• Notes: independent community experiment built on the MOSS Image architecture




![Nix Comic [Multiple Styles]](/media/civitai--2836231-nix-comic-multiple-styles/cover-320.webp)
![Nix Comic [Multiple Styles]](/media/civitai--2836231-nix-comic-multiple-styles/cover.webp)




Comments