# Animagine XL 3.1 Guide to Anime Prompts, Settings and Setup > Published 2026-04-09, updated 2026-09-06 · https://www.promptzone.com/karan_reddy/animagine-xl-anime-ai-image-generator-37al --- title: Animagine XL 3.1 Guide to Anime Prompts, Settings and Setup published: true description: Learn how to run Animagine XL 3.1, structure anime prompts, choose sampling settings, and check the checkpoint's license. tags: ai,stablediffusion,imagegeneration --- Animagine XL 3.1 is Cagliostro Research Lab's anime image checkpoint built on Stable Diffusion XL. Download its weights from the official Hugging Face repository and run it with an SDXL-compatible pipeline. [Model card][card] This guide focuses on a repeatable prompting workflow for version 3.1: establish a baseline, change one description at a time, and inspect the result before adding more controls. ## What are the key facts about Animagine XL 3.1? | Field | Verified detail | | --- | --- | | Developer | Cagliostro Research Lab, also presented as Cagliostro Labs. [Model card][card] | | Released | March 18, 2024, for version 3.1. [Announcement][release] | | Type | Anime text-to-image checkpoint based on SDXL and continued from Animagine XL 3.0. [Announcement][release] | | Size or parameters | `animagine-xl-3.1.safetensors`: 6.94 GB, as listed in the official repository; file size is not a VRAM requirement. [Model files][files] | | License and access | Downloadable weights; the current card specifies CreativeML Open RAIL++-M. [Model card][card], [license][license] | | Where it runs | A compatible Python environment using Diffusers; the card demonstrates CUDA inference. [Model card][card] | ## What can you control with Animagine XL 3.1 prompts? The release emphasizes anime character knowledge and ordered tags. Its authors expanded the represented franchises and added aesthetic controls, making version 3.1 useful for character illustration experiments. [Announcement][release] The practical attraction is explicit control over the description. You can name a character, clothing, framing, expression, and background separately, then decide which of those elements deserves the next revision. Use a small character brief before generating. Write down the features that must survive: hair shape, garment color, expression, and whether the image needs empty space for a layout. Treat each output as a candidate. A picture that matches the palette but misses the expression may be useful for color exploration while remaining unsuitable for the final illustration. The authors provide quality, aesthetic, and era tags as distinct controls. These are training-related prompt conventions; they should not be read as measured quality scores for the image you receive. [Announcement][release] For example, test a modern treatment and a vintage treatment using the same subject description. Judge line work, shading, and composition separately instead of choosing whichever result happens to contain more detail. ## What are the limits and license terms of Animagine XL 3.1? The model card identifies an anime focus, a preference for Danbooru-style tags, and continuing problems with hands and anatomy. It also cautions that short prompts may not supply enough detail. [Model card][card] Plan a review pass around those limitations. Enlarge the image to inspect fingers, overlapping arms, and small accessories, then view it at the intended display size to judge expression and silhouette. A useful revision names the observed problem. If a sleeve hides the hand awkwardly, first simplify the pose or framing; avoid changing the background, outfit, and sampling settings in the same attempt. Licensing also depends on the artifact you obtain. The launch announcement describes Fair AI Public License 1.0-SD, while the current repository explicitly replaces earlier community tags with Open RAIL++-M. [Announcement][release], [model card][card] Save the license associated with your downloaded revision. The linked Open RAIL++-M agreement contains use restrictions and redistribution conditions, so keep it with any checkpoint inventory you maintain. [License][license] ## How do you install and prompt Animagine XL 3.1? ### Establish a working Python baseline Install a PyTorch build suitable for your hardware, then install `diffusers`, `transformers`, `accelerate`, and `safetensors`. The official card documents this package set and a CUDA example. [Model card][card] The following example uses the published checkpoint with an original character prompt. Run it in an environment with a working CUDA-enabled PyTorch installation and enough memory for the pipeline. ```python import torch from diffusers import DiffusionPipeline, EulerAncestralDiscreteScheduler pipe = DiffusionPipeline.from_pretrained( "cagliostrolab/animagine-xl-3.1", torch_dtype=torch.float16, use_safetensors=True, ) pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config) pipe.to("cuda") image = pipe( "1girl, solo, adult, short blue hair, yellow raincoat, " "standing, train platform, evening, best quality, very aesthetic", negative_prompt="low quality, text, watermark, extra digits", width=832, height=1216, guidance_scale=6, num_inference_steps=28, generator=torch.Generator("cuda").manual_seed(42), ).images[0] image.save("anime-study.png") ``` The dimensions, guidance range, step count, and Euler Ancestral sampler follow the card's recommendations. Scheduler replacement uses the Diffusers scheduler interface. [Model card][card], [scheduler documentation][scheduler] ### Build the prompt in deliberate passes Begin with the subject. For a known character, the authors recommend putting the character and franchise near the start; for an original design, use concrete visual attributes without inventing a franchise tag. [Announcement][release] Next, choose the composition: a portrait, an upper-body study, or a full scene. Prefer one clear framing intention so you can judge whether the subject occupies the space you need. Add the setting and light after the character brief is readable. Keep a notebook with the prompt change, selected output, and the reason you preferred it; this makes your choices easier to revisit. Try the checkpoint by itself before adding adapters. The [PromptZone LoRA guide](/tara_suzuki/how-to-use-loras-in-comfyui-in-2026-load-stack-and-troubleshoot-235e) provides a separate workflow for evaluating those additions. If memory becomes a problem, consult the SDXL documentation's offloading options. It documents moving components between CPU and GPU; it does not make a universal promise for every hardware configuration. [SDXL documentation][sdxldocs] ## How does Animagine XL 3.1 compare with SDXL base? | Option | Suitable starting point | | --- | --- | | Animagine XL 3.1 | Anime character studies using the checkpoint's tag conventions. [Model card][card] | | Stable Diffusion XL base 1.0 | General image generation with a base checkpoint that can run independently of its optional refiner. [SDXL card][sdxl] | Use the [SDXL model guide](/tara_suzuki/best-sdxl-models-in-2026-realistic-anime-and-all-purpose-checkpoints-116) to organize further comparisons around your intended style. For a useful comparison, keep the character brief and output shape consistent. Allow each checkpoint its documented settings, and record the time you spend repairing the result as part of your own evaluation. ## What else should you know about Animagine XL 3.1? ### Is Animagine XL 3.1 a LoRA? Animagine XL 3.1 is a full SDXL-derived checkpoint. Load it as the main image model; its official Diffusers example downloads the pipeline from the Cagliostro repository. [Model card][card] ### Can Animagine XL create realistic photographs? Animagine XL 3.1 is designed for anime images; its authors say it is unsuitable for realistic photos. Choose a photography-focused model for that task, and evaluate it using photographic criteria. [Model card][card] ### Which license applies to the current download? The current Animagine XL 3.1 model card specifies CreativeML Open RAIL++-M and says it supersedes earlier community license tags. Check the license for your downloaded revision, especially for independently redistributed copies. [Model card][card], [license][license] ### Should I start by increasing the sampling steps? For Animagine XL 3.1, the model card recommends Euler Ancestral, a CFG scale around 5–7, and fewer than 30 sampling steps. Start there and revise the description when the content is wrong. Keep a fixed seed during a small comparison so you can more clearly judge the effect of the change. [Model card][card] ## Sources - [Animagine XL 3.1 model card][card] - [Animagine XL 3.1 checkpoint files][files] - [Cagliostro release announcement][release] - [CreativeML Open RAIL++-M license][license] - [Diffusers scheduler guide][scheduler] - [Diffusers SDXL guide][sdxldocs] - [Stable Diffusion XL base model card][sdxl] [files]: https://huggingface.co/cagliostrolab/animagine-xl-3.1/tree/main [card]: https://huggingface.co/cagliostrolab/animagine-xl-3.1 [release]: https://cagliostrolab.net/posts/animagine-xl-v31-release [license]: https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/blob/main/LICENSE.md [scheduler]: https://huggingface.co/docs/diffusers/en/using-diffusers/schedulers [sdxldocs]: https://huggingface.co/docs/diffusers/v0.35.1/en/using-diffusers/sdxl [sdxl]: https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0 ## Related guides on PromptZone - [Best SDXL Models in 2026](/tara_suzuki/best-sdxl-models-in-2026-realistic-anime-and-all-purpose-checkpoints-116) - [ComfyUI 2026: The Complete Guide](/tomas_novak/comfyui-2026-the-complete-guide-to-power-user-ai-image-generation-1g17) - [AI Model Releases Timeline](/ai-model-releases)