<p>Anima is a 2 billion parameter text-to-image model created via a collaboration between CircleStone Labs and Comfy Org. It is focused mainly on anime concepts, characters, and styles, but is also capable of generating a wide variety of other non-photorealistic content. The model is designed for making illustrations and artistic images, and will not work well at realism.</p><p>It is trained on several million anime images and about 800k non-anime artistic images. No synthetic data was used for training. The knowledge cut-off date for the anime training data is September 2025.</p><h1 id="versions">Versions</h1><ul><li><p>Anima-Base</p><ul><li><p>The pretrained, unrefined base model. Maximum flexibility, diversity, and style adherence.</p></li><li><p>LoRAs should be trained using this version.</p></li></ul></li><li><p>Anima-Aesthetic</p><ul><li><p>Fine-tuned for better consistency and a higher quality default art style.</p></li></ul></li><li><p>Anima-Turbo</p><ul><li><p>Distilled version for fast generations.</p></li><li><p>Use at CFG 1 and 8-12 steps.</p></li><li><p>The distillation process also increases stability and gives the model a strong default style, but reduces diversity.</p></li></ul></li></ul><p>I recommend starting with Anima-Turbo. On average, it is only slightly worse than Anima-Aesthetic, while being very fast to generate (and much cheaper if you use it on an online platform that scales the cost with step count). This makes it very convenient for quickly iterating on prompts. The increased stability can even make it better than Aesthetic in some cases.</p><h1 id="installing-and-running">Installing and running</h1><p>Get the text encoder and VAE from the <a target="_blank" rel="ugc" href="https://huggingface.co/circlestone-labs/Anima">HuggingFage page</a>.</p><p>The model is natively supported in ComfyUI. The model files go in their respective folders inside your model directory:</p><ul><li><p>anima-base-v1.0.safetensors goes in ComfyUI/models/diffusion_models</p></li><li><p>qwen_3_06b_base.safetensors goes in ComfyUI/models/text_encoders</p></li><li><p>qwen_image_vae.safetensors goes in ComfyUI/models/vae (this is the Qwen-Image VAE, you might already have it)</p></li></ul><h2 id="generation-settings">Generation settings</h2><ul><li><p>Works at resolutions between 512^2 and 1536^2 pixels.</p></li><li><p>30-50 steps, CFG 4-6.</p><ul><li><p>The Aesthetic version can tolerate lower CFGs such as 3, and often looks better with them.</p></li></ul></li><li><p>A variety of samplers work. Some of my favorites:</p><ul><li><p>er_sde: neutral style, flat colors, sharp lines. I use this as a reasonable default.</p></li><li><p>euler_a: Softer, thinner lines. Can sometimes tend towards a 2.5D look. CFG can be pushed a bit higher than other samplers without burning the image.</p></li><li><p>dpmpp_2m_sde_gpu: similar in style to er_sde but can produce more variety and be more "creative". Depending on the prompt it can get too wild sometimes.</p></li><li><p>euler: a basic sampler that is a bit more creative than er_sde. Good with the Turbo and Aesthetic versions, since those are naturally more stable.</p></li></ul></li><li><p>If going for a more realistic / painterly look, the beta57 scheduler (ComfyUI RES4LYF custom node pack) can help make better textures, since it puts more emphasis on low-noise timesteps.</p></li></ul><h1 id="prompting">Prompting</h1><p>The model is trained on Danbooru-style tags, natural language captions, and combinations of tags and captions.</p><ul><li><p>Use lowercase for tags, and spaces instead of underscores. Score tags are the only tags that use underscores.</p></li><li><p>Recommended positive prefix: "masterpiece, best quality, score_7, safe, "</p></li><li><p>Recommended negative: "worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, chromatic aberration"</p></li><li><p>When using a tag that is different between Danbooru and Gelbooru, prefer the Gelbooru version.</p></li><li><p>Prompt weighting works, but needs a weight higher than typically used for SDXL. Example: "(chibi:2)"</p></li></ul><h2 id="aesthetic-version-prompting">Aesthetic Version Prompting</h2><p>Anima-Aesthetic is fine-tuned only on high quality images, with all of the quality tags stripped out from the captions. You don't need to use quality tags in the positive at all, but "masterpiece, best quality, " is safe to leave in. If you are getting too much noise, artifacts, details, or "slop", try some combination of: 1) lower the CFG, 2) remove score_* tags from the positive and negative, 3) add "anime coloring" to the prompt (maybe with a strength less than 1).</p><h2 id="tag-order">Tag order</h2><p>[quality/meta/year/safety tags] [1girl/1boy/1other etc] [character] [series] [artist] [general tags]</p><p>Within each tag section, the tags can be in arbitrary order.</p><h2 id="quality-tags">Quality tags</h2><p>Human score based: masterpiece, best quality, good quality, normal quality, low quality, worst quality</p><p>PonyV7 aesthetic model based: score_9, score_8, ..., score_1</p><p>You can use either the human score quality tags, the aesthetic model tags, both together, or neither. All combinations work.</p><h2 id="time-period-tags">Time period tags</h2><p>Specific year: year 2025, year 2024, ...</p><p>Period: newest, recent, mid, early, old</p><h2 id="meta-tags">Meta tags</h2><p>highres, absurdres, anime screenshot, jpeg artifacts, official art, etc</p><h2 id="safety-tags">Safety tags</h2><p>safe, sensitive, nsfw, explicit</p><h2 id="artist-tags">Artist tags</h2><p>Prefix artist with @. E.g. "@big chungus". <strong>You must put @ in front of the artist.</strong> The effect will be very weak if you don't.</p><h2 id="full-tag-example">Full tag example</h2><p>year 2025, newest, normal quality, score_5, highres, safe, 1girl, oomuro sakurako, yuru yuri, @nnn yryr, smile, brown hair, hat, solo, fur-trimmed gloves, open mouth, long hair, gift box, fang, skirt, red gloves, blunt bangs, gloves, one eye closed, shirt, brown eyes, santa costume, red hat, skin fang, twitter username, white background, holding bag, fur trim, simple background, brown skirt, bag, gift bag, looking at viewer, santa hat, ;d, red shirt, box, gift, fur-trimmed headwear, holding, red capelet, holding box, capelet</p><h2 id="tag-dropout">Tag dropout</h2><p>The model was trained with random tag dropout. You don't need to include every single relevant tag for the image.</p><h2 id="dataset-tags">Dataset tags</h2><p>To improve style and content diversity, the model was additionally trained on two non-anime datasets: LAION-POP (specifically the ye-pop version) and DeviantArt. Both were filtered to exclude photos. Because these datasets are qualitatively different from anime datasets, captions from them have been labeled with a "dataset tag". This occurs at the very beginning of a prompt followed by a newline. Optionally, the second line can contain either the image alt-text (ye-pop) or the title of the work (DeviantArt). Examples:</p><pre><code>ye-pop
For Sale: Others by Arun Prem
Abstract, oil painting of three faceless, blue-skinned figures. Left: white, draped figure; center: yellow-shirted, dark-haired figure; right: red-veiled, dark-haired figure carrying another. Bold, textured colors, minimalist style.</code></pre><pre><code>deviantart
Flame
Digital painting of a fiery dragon with glowing yellow eyes, black horns, and a long, sinuous tail, perched on a glowing, molten rock formation. The background is a gradient of dark purple to orange.</code></pre><h2 id="natural-language-prompting-tips">Natural language prompting tips</h2><ul><li><p>Follow standard English capitalization rules for character and series names.</p></li><li><p>If using pure natural language, more descriptive is better. Aim for at least 2 sentences. Extremely short prompts can give unexpected results.</p></li><li><p>You can mix tags and natural language in arbitrary order.</p></li><li><p>You can put quality / artist tags at the beginning of a natural language prompt.</p><ul><li><p>"masterpiece, best quality, @big chungus. An anime girl with medium-length blonde hair is..."</p></li></ul></li><li><p>Name a character, then describe their basic appearance.</p><ul><li><p>"Digital artwork of Fern from Sousou no Frieren, with long purple hair and purple eyes, wearing a black coat over a white dress with puffy sleeves..."</p></li><li><p>This is extra important when prompting for multiple characters. If you just list off character names with no description of appearance, the model can get confused.</p></li></ul></li></ul><h1 id="limitations">Limitations</h1><ul><li><p>The model doesn't do realism well. This is intended. It is an anime / illustration / art focused model.</p></li><li><p>The model may generate undesired content, especially if the prompt is short or lacking details.</p><ul><li><p>Avoid this by using the appropriate safety tags in the positive and negative prompts, and by writing sufficiently detailed prompts.</p></li></ul></li><li><p>The model isn't great at text rendering. It can generally do single words and sometimes short phrases, but lengthy text rendering won't work well.</p></li><li><p>The base version is a true base model. It hasn't been aesthetic tuned on a curated dataset. The default style is very plain and neutral, which is especially apparent if you don't use artist or quality tags.</p></li></ul><h1 id="finetuning-tips">Finetuning tips</h1><ul><li><p><strong>Don't train the LLM adapter.</strong> My own training script, diffusion-pipe, lets you set llm_adapter_lr=0 to completely disable training it, and the example config has this as a default.</p><ul><li><p>Other trainers like sd-scripts have similar options that should be used.</p></li><li><p>The LLM adapter processes the text embeddings before they get to the diffusion model, and therefore has an outsized influence on the generated images. The adapter itself contains a surprising amount of knowledge and is easy to degrade by training it.</p></li></ul></li><li><p>Use a low learning rate. For a rank 32 LoRA, start with 2e-5 and adjust up or down from there.</p><ul><li><p>As a base model, there is no aggressive aesthetic tuning or RLHF you need to overcome when finetuning.</p></li><li><p>The model has an extremely large and diverse amount of visual concepts baked in already. A light touch is all you need.</p></li></ul></li><li><p><a target="_blank" rel="ugc" href="https://civarchive.com/models/2536147">Example</a> of a style LoRA, with dataset and configs shared.</p></li></ul><h1 id="license">License</h1><p>This model is licensed under the CircleStone Labs Non-Commercial License. The model and derivatives are only usable for non-commercial purposes. Additionally, this model constitutes a "Derivative Model" of Cosmos-Predict2-2B-Text2Image, and therefore is subject to the <a target="_blank" rel="ugc" href="https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/">NVIDIA Open Model License Agreement</a> insofar as it applies to Derivative Models.</p><p>If you would like a commercial license, please email tdrussell@circlestone.ai</p><p>Built on NVIDIA Cosmos.</p>

Anima

<p>Alternate version of Anima-Aesthetic. This one consists only of an aesthetic finetune on 10k images, with none of the additional style adjustment and stabilization loras merged in like 1.0. I personally like it less, and think it tends toward a 2.5D, shiny, plastic, AI-slop look. But maybe some people prefer it.</p>

Files with SHA256: 49d1e23b9a7415ac56f4719017570a5f8d295184faba578e07fe76f41bee0fcc