Animagine XL 3.1
Image Available Full comparison ↗ cagliostrolab/animagine-xl-3.1 · by cagliostrolab
· Latent diffusion
· Built on Stable Diffusion XL 1.0
Pricing — 1 offering(s)
About this offering: A fine-tune of the Animagine XL V3 series (SDXL-based), specialised for anime-style images. Open weights under the CreativeML Open RAIL++-M licence. This listing is a third-party host (Replicate), not cagliostrolab itself. Cagliostrolab has continued the Animagine line past 3.1, so this is treated as a previous-generation model.
Default
- $0.000975 / second
Current
2026-10-18 → present · Nvidia L40S
Billed per second of GPU time on an Nvidia L40S (community-hosted wrapper, not a cagliostrolab endpoint). A…
Billed per second of GPU time on an Nvidia L40S (community-hosted wrapper, not a cagliostrolab endpoint). A typical image takes about 4 seconds, roughly $0.0038 per image (Replicate's median run cost; an estimate that varies with inputs).
Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.
See how Animagine XL 3.1 fits into a cost-aware routing setup
See how →About Animagine XL 3.1
Animagine XL 3.1 is a specialist anime-illustration model from cagliostrolab, a community group founded in 2023. It is not a from-scratch architecture: it is a fine-tune on Stable Diffusion XL's U-Net backbone, and a direct continuation of the Animagine XL V3 series — built on top of the 3.0 checkpoint by incremental learning rather than retrained from SDXL. Per the model card it was trained on roughly 870,000 ordered and tagged anime images across a pretraining stage and two fine-tuning stages, using the Danbooru-style tagging convention.
It is purpose-built for anime and illustration output and is not a general-purpose alternative to FLUX or Stable Diffusion 3.5 — photorealism is explicitly not a goal. Getting good prompt adherence requires the tag-based (Danbooru-style) prompting the model was trained on, not natural-language prose. It runs fast (recommended under 30 sampling steps; about 4 seconds on an L40S on Replicate) at sub-cent cost per image, and the weights are open under the CreativeML Open RAIL++-M licence — commercial use is permitted — with an active community ecosystem of LoRAs and further fine-tunes.
Two caveats carry from the model card. Hand and anatomy rendering remain imperfect despite being an explicit focus of this version. And it can generate NSFW content unprompted despite mitigation efforts, so it should be deployed with output filtering. There are no independent GenEval or T2I-CompBench scores for it — the capability ratings in this entry are a qualitative read of the model card and community reception, and Modelglass records the confidence as low.
Capability profile
How Animagine XL 3.1 rates across core capability dimensions, with the task-level evidence behind each rating.
Purpose-built and heavily specialized for anime-style illustration.
Not a goal of this model — anime-only stylistic range.
Strong within its tag-based prompting convention; weaker on free-form natural-language prompts (model card recommends Danbooru-style tags, not prose).
Model card cites improved hand anatomy vs. earlier Animagine versions, but anatomy/hands remain an acknowledged weak point.
Native SDXL-class resolutions plus several extended aspect ratios (e.g. 832x1216, 1536x640) per the model card.
Recommended under 30 sampling steps (Euler Ancestral); Replicate's hosted version completes in ~4s on an L40S.
Ratings are estimated — limited independent data is available for this model.
Operator guidance
Route here specifically for anime/illustration-style output; it is not a general-purpose photorealistic alternative to FLUX or Stable Diffusion 3.5. Prompting must follow the tag-based (Danbooru-style) convention the model was trained on, not natural-language prompts, to get good adherence.
Use cases
- Anime/illustration-style image generation where photorealistic models are the wrong fit
- Cost-sensitive, high-volume anime-style generation (sub-cent per image on Replicate)
- Local/self-hosted anime generation (open weights, active community LoRA/fine-tune ecosystem)
Provenance notice
- Open weights; RAIL++-M license permits commercial use while prohibiting illegal/harmful-content use
- Model card explicitly warns it may generate NSFW content unprompted despite mitigation — deploy with output filtering
Limitations
- Anime-style only — not competitive on photorealism or general-purpose generation
- Requires tag-based, not natural-language, prompting for reliable prompt adherence
- Hand/anatomy rendering imperfect, despite being an explicit improvement focus for this version
- Model card notes it may generate NSFW content unprompted despite mitigation efforts — deploy with output filtering
- No independent GenEval/T2I-CompBench scores tracked yet; capability ratings below are qualitative, not benchmark-derived