Mugen - Modernized Anime SDXL Base, or how to make Bluvoll tiny bit less sane
New AI image model diverges from prior NoobAI lineage with a focus on character fidelity.
Cabal Research, a group including developers Anzhc and Bluvoll, has launched Mugen, a new base model for the Stable Diffusion XL (SDXL) framework specifically tuned for anime-style art. The release marks a formal break from the team's previous, more cryptically named "NoobAI" model lineage, opting for a cleaner branding. Mugen is described as the continuation of an experimental pipeline focused on integrating the Flux 2 VAE (Variational Autoencoder), a component critical for image detail and quality, into the SDXL architecture.
A core technical priority for this iteration was improving what the developers term "character knowledge"—the model's ability to consistently generate and recognize specific character traits, poses, and styles. To quantify progress in this niche area, the team developed a specialized benchmark, moving beyond generic image quality metrics. The model is now available for download on the Hugging Face platform, allowing other AI artists and researchers to test its capabilities. The announcement humorously notes the project's toll on team member Bluvoll's sanity, highlighting the intense effort behind refining these open-source AI art tools.
- Mugen is a new SDXL base model from Cabal Research, focused on anime-style generation.
- It continues experiments with the Flux 2 VAE architecture, diverging from the prior 'NoobAI' model series.
- Development prioritized 'character knowledge,' with a custom benchmark created to measure gains in this area.
Why It Matters
Provides the open-source AI art community with a specialized, high-fidelity tool for consistent anime character generation.