Mugen by Cabal Research Elevates Anime Character Creation

Mugen is a new AI image model that converts SDXL architecture to Flux 2 VAE, designed specifically for anime-style generation. It represents a significant departure from the original NoobAI models after seven additional epochs of training.
Developed by Cabal Research, this project cost under $8,000 to complete and focuses on preserving character knowledge while improving texture and pattern quality. The team created an in-house benchmarking system to measure character similarity across 1,815 characters, including gachas and vtubers.
Model Size: 7GB VRAM GPU: requirements vary
What Mugen Offers
- Base model designed for standard-friendly local training.
- Three additional variants: Aesthetic, and two opinionated dataset tunes.
- Improved texture and pattern generation impossible with SDXL VAE.
- Custom character discrimination model trained on 1.2 million images.
- Native support for ComfyUI and A1111 WebUI (ReForge).
Anime artists and hobbyists creating character-focused content will find the improved character adherence useful. The model handles series-specific characters better when combined with appearance tags, making it practical for fan art projects or original character development.
Development notes and practical tips
The team notes a brown color bias inherited from Flux 2 VAE, which users can correct by adding sepia or brown themes to negative prompts. For best results with characters, include series or game names alongside character triggers. The developers recommend specific sampling settings: 20-28 steps, CFG between 4-7, and shift values of 8-12.
'We prioritized keeping our training as standard-friendly as possible, so local community can easily train on it like on a new Base Model, which it practically is.'
ReForge users should set preview method to 'Approx Cheap' for proper visualization, as high-quality preview methods do not currently support Flux2VAE.
Download Mugen on Hugging Face.