Animagine XL 4.0: tag order, score and year tags, and recommended settings

Animagine XL 4.0 (Cagliostro Research Lab) is an anime finetune retrained from SDXL 1.0 on about 8.4 million images (knowledge cut-off 7 January 2025) with tag-ordered captions; the 4.0 Opt update of 13 February 2025 reports better stability, anatomy, noise, saturation and colour accuracy. It runs in ComfyUI, the Stable Diffusion WebUI and diffusers. Prompting: use tags, not natural language, in this order: 1girl/1boy/1other, character name, series, rating, then anything else, and end with the quality tags masterpiece, high score, great score, absurdres. Recommended negative prompt: lowres, bad anatomy, bad hands, text, error, missing finger, extra digits, fewer digits, cropped, worst quality, low quality, low score, bad score, average score, signature, watermark, username, blurry. Special tags: quality tags (masterpiece, best quality, low quality, worst quality); score tags (high, great, good, average, bad, low score), which the authors say steer quality more strongly; year tags from year 2005 to year 2025 for era-specific style; and rating tags (safe, sensitive, nsfw, explicit), which the card asks to be used responsibly and lawfully. Settings: CFG 4 to 7 (5 recommended), 25 to 28 steps (28 recommended), Euler Ancestral. Sizes: 1024x1024; landscape 1152x896, 1216x832, 1344x768, 1536x640; portrait 896x1152, 832x1216, 768x1344, 640x1536. Higher resolutions such as 1536x1536 may degrade. diffusers: StableDiffusionXLPipeline.from_pretrained("cagliostrolab/animagine-xl-4.0", torch_dtype=torch.float16, use_safetensors=True, custom_pipeline="lpw_stable_diffusion_xl", add_watermarker=False); the weights are already fp16, so no variant argument is needed, and the lpw pipeline handles long weighted prompts. The example uses 832x1216, guidance_scale=5 and 28 steps. Note that custom_pipeline loads community pipeline code. Limitations noted by the card: weak hands and finger counts, no text rendering, lower accuracy for recent characters, multi-character scenes need careful prompting, and style may need explicit style tags. Licence: unmodified CreativeML Open RAIL++-M; the card summarises it as permitting commercial use and modification with use restrictions, and points to the SDXL licence as authoritative. Based on the cagliostrolab/animagine-xl-4.0 card pinned to revision 2b7c1b397761bf5bd3cc42e5b39ec99314a75a96 and observed on 2026-10-27. Nothing was run here; the sample images were not inspected, and the card does not say which file holds the 4.0 Opt weights. The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.

{
  "citations": [
    {
      "url": "https://huggingface.co/cagliostrolab/animagine-xl-4.0/raw/2b7c1b397761bf5bd3cc42e5b39ec99314a75a96/README.md",
      "note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 5215:5342, 219:1082, 5746:5878, 5828:5878, 6257:6385, 7777:7960, 14413:14660; exact excerpt SHA256 ad4d94b613456b643b9695996373b0db0a56ba100b205f1581d2823b178b5ac8,313545c8a580835457db560cba06664187e380c9b12c2b8215e219b77edd6ebc,53ae878f40339b32ac301c173e04538b0f3b36d18e89562f7"
    }
  ]
}