MosYume
I. Introduction
MosYume is a text-to-image model fine-tuned from MOSS-Image-Base, a high-quality image generation foundation model developed by OpenMOSS-Team.
This is an independent, community-driven experiment created out of a passion for exploring the untapped potential of the MOSS architecture. The primary goal of this fine-tune is to resolve anatomical inconsistencies and stabilize irregular weight distributions present in the original base model. (Note: This project is independently developed and is not affiliated with any research lab or company).
Version v0.1 Highlights
- Pure Human-Created Dataset: Trained on 20 million curated images spanning both anime and realistic styles, without any AI-generated images or model distillation.
- Balanced Domain Distribution: Maintains 1:1 ratio between anime and realistic imagery.
- Multi-Resolution Training: Trained across dynamic resolutions ranging from 1024px to 2048px.
- Pure Base Checkpoint: Released as a clean, base model without Supervised Fine-Tuning (SFT), making it ideal for further downstream training or LoRA creation.
- Anime Data Sources: Sourced from Danbooru (updated up to September 4, 2026), e621, Gelbooru, Pixiv, and Konachan (with other datasets were updated at least july 30, 2026).
Key Features
- High-Quality Generation: Produces crisp, detailed anime illustrations featuring sharp linework, vibrant colors, and smooth shading.
- Superior Character & Tag Comprehension: Deeply understands Danbooru tags and character concepts, delivering accurate and coherent character representations.
- Intricate Fine Details: Renders complex accessories, fabric textures, hairstyles, and rich backgrounds with high visual clarity.
- Native 2K Support: Capable of native high-resolution inference up to 2048×2048.
II. Model Architecture & Components
- Text Encoder: Pre-trained Qwen3.5-9B
- Variational Autoencoder (VAE): Pre-trained Flux.2 VAE
- Image Backbone: Fine-tuned MOSS-Image-Base Backbone
III. Recommended Inference Settings
For optimal generation quality, use the following parameters:
- CFG Scale:
3 – 5 - Sampling Steps:
30 - Sampler:
er_sde - Scheduler:
simple
IV. Acknowledgments
Special thanks to the open-source community for making this project possible:
- narugo1992 – For providing and maintaining the invaluable Danbooru dataset.
- OpenMOSS-Team – For open-sourcing the impressive MOSS-Image-Base model.
- Neta.art – For sharing the richly detailed dataset originally curated for Neta Lumina.
