Animagine XL

Cagliostro Research Lab builds its Animagine models as full fine-tunes of Stable Diffusion XL, and version 3.1, the release this page links, continued training from 3.0 on roughly 870,000 tagged images to add more franchise characters, steadier hands and a cleaner tag scheme. Prompts work best as Danbooru-style tags in a set order: subject count, character, series, then everything else. Special tokens push results toward a quality tier, a content rating, an era of anime art and an aesthetic score, and the card suggests CFG 5 to 7, fewer than 30 steps and the Euler a sampler at SDXL resolutions near one megapixel. A single safetensors checkpoint loads in ComfyUI or the Automatic1111 web UI, with a Diffusers layout too, so any SDXL-capable GPU runs it. Animagine XL 4.0, retrained from SDXL base on about 8.4 million images, is the better start for new projects. Licensing for 3.1 has moved to CreativeML Open RAIL++-M, replacing an earlier Fair AI Public License tag, and allows commercial use within the usual use restrictions. Photorealism and plain-English prompts are weak spots, and it can return NSFW images unprompted.