Animagine XL is an anime-style-focused Stable Diffusion XL model series developed by Cagliostro Research Lab. It is suitable for anime creators who want to precisely control the image using tags, with a prompt style close to Danbooru tags, and can guide the generation effect with special tags such as quality, score, year, and rating.
Main capabilities and training information: The current mainline version is 4.0 (also called Anim4gine), retrained from SDXL 1.0 (not merely layered on top of 3.x), using about 8.4 million anime-style images, with a knowledge cutoff around early January 2025. The public page lists its license as CreativeML Open RAIL++-M, in the same series as the SDXL terms; the full license should be read before commercial use.
Version history: 3.0 (January 2024) fine-tuned hand and concept understanding and adopted NovelAI-style tag ordering habits; 3.1 (March 2024) added character and series knowledge and aesthetic tags, and reorganized quality/rating/year tags; 4.0 / Anim4gine (January 2025) was retrained from SDXL 1.0, with about 8.4 million images and about 2650 GPU hours, and is officially described as more artistic and more prompt-following than 3.1; among 4.0 Opt and Zero (February 2025), Opt provides additional refinement for everyday generation, while Zero is a base model for LoRA and further fine-tuning.
How to use: Weights can be loaded in local tools such as ComfyUI, Forge, or Automatic1111. The starting recipe given in the official 4.0 guide includes: arrange in the order of "subject count, character, series, rating, other tags," with quality tags such as masterpiece, high score, great score, and absurdres placed at the end; sampling suggestions are Euler a, about 25–28 steps, CFG about 4–7 (commonly 5), and standard SDXL resolution; when invoking a character, also add the series or copyright tag, because characters were paired with their series during training. You can also directly open the official Playground to try it in the browser and switch between version tags such as v4.0, v3.1, and v3.0, so you can start without a local GPU installation; demoThe page may take a moment to load.
Use cases: image generation of anime characters and recent series works, creative workflows that want to control composition with tags, and use as a base model for LoRA training or further fine-tuning (the Zero variant).
Limitations: The above content is based on the official public introduction; specific performance benchmarks, runtime hardware requirements, or pricing information are not provided; for commercial use, you should consult the full license text yourself.
