Anima
A 2B anime and illustration image model from CircleStone Labs and Comfy Org, with a Turbo version for 8 to 12 steps. Non-commercial weights.
In short. Anima is a small AI image model for anime, illustrations and other non-photorealistic art. It runs on an ordinary gaming GPU, and its Turbo version needs only 8 to 12 steps per picture. The weights are for non-commercial use only, but the pictures you make with them can be used commercially. Hosting the model in a paid service needs a separate license.
Summary
Anima is a 2-billion-parameter text-to-image model made by CircleStone Labs together with Comfy Org. It focuses on anime concepts, characters and styles and also handles other non-photorealistic art, but the card says plainly that it is not meant for realism. It was trained on several million anime images and about 800,000 non-anime artistic images, with no synthetic data, and the anime data stops at September 2025. Preview checkpoints appeared in January 2026, the Base v1.0 checkpoint on 14 May 2026, and the Aesthetic and Turbo versions in July and August; the repository has about 2,600 likes.
There are three kinds of checkpoint. Base is the pretrained, unrefined model with the most flexibility and diversity, and the developer says LoRAs should be trained on it. Aesthetic is fine-tuned for a more consistent, higher-quality default style, in versions 1.0, 1.0b and 1.1. Turbo is distilled for fast generation at a guidance scale of 1 and 8 to 12 steps, which makes it more stable but less diverse; the developer suggests starting with Turbo because it is only slightly worse than Aesthetic on average. The other versions use 30 to 50 steps at a guidance scale of 4 to 5, and the model works between 512² and 1536² pixels. Prompts can be Danbooru-style tags, natural-language captions or a mix, with quality, year and safety tags and artist names written with an @ prefix.
Each checkpoint is a 4.18 GB BF16 file of about 2.09 billion parameters in ComfyUI's format. It pairs with a 1.19 GB Qwen text encoder and the 0.25 GB Qwen-Image VAE, so the whole pipeline is about 5.6 GB; the model card says an adapter processes the text embeddings before they reach the diffusion model, and advises against training it. ComfyUI supports Anima natively, and the card lists CivitAI, TensorArt and other hosted platforms that offer it.
The license is the CircleStone Labs Non-Commercial License v1.2. Because the model is a derivative of NVIDIA's Cosmos-Predict2-2B-Text2Image, the NVIDIA Open Model License applies as well, as far as it covers derivative models. The restriction applies to the model and not to its outputs: selling images, taking paid commissions and using pictures as assets in a paid game are allowed, and an individual may sell derivative weights. Hosting the model behind a paid API or paid generation site, or building it into a monetized product, needs a separate commercial license from CircleStone Labs. The card publishes no benchmark scores, and it lists limits: poor realism, weak long-text rendering, and undesired content from short prompts unless you use safety tags.
Good for
- Anime, manga-style and illustration work on an ordinary gaming GPU, including characters and artist styles written as tags
- Fast prompt iteration with the Turbo version at 8 to 12 steps, or higher-quality output with Aesthetic
- Training style and character LoRAs on the Base model, following the card's tips to use a low learning rate and leave the text adapter alone
- ComfyUI workflows, since support is built in and the developer ships a model-comparison workflow
Watch out for
- The weights are non-commercial, so hosting the model in a paid service or inside a monetized product needs a separate license, although the images you generate can be sold
- NVIDIA's Open Model License applies as well, because Anima is a Cosmos derivative
- It is not for photorealism, and it handles single words and short phrases of text but not long text
- Short or vague prompts can produce undesired content, so use the safety tags and write at least two sentences when you use plain language
- The card gives no benchmark scores, so judge quality by trying it
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of Anima. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| Anima-Turbo v1.1 (BF16) BF16 | 2.09B | 5.6 GB | 6.44 GB | Laptop or phone |
| Anima-Aesthetic v1.1 (BF16) BF16 | 2.09B | 5.6 GB | 6.44 GB | Laptop or phone |
| Anima-Base v1.0 (BF16) BF16 | 2.09B | 5.6 GB | 6.44 GB | Laptop or phone |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
The license for CircleStone Labs' Anima image models. The weights are free only for non-commercial use, but the images you generate can be used commercially. Hosting the model for payment, or building it into a monetized product, needs a separate commercial license from CircleStone Labs.
- Attribution or model name must be shown
- Acceptable-use policy applies
More about CircleStone Non-Commercial
See also
Frequently asked questions
- What is Anima?
- Anima is a small AI image model for anime, illustrations and other non-photorealistic art. It runs on an ordinary gaming GPU, and its Turbo version needs only 8 to 12 steps per picture. The weights are for non-commercial use only, but the pictures you make with them can be used commercially. Hosting the model in a paid service needs a separate license.
- Can I use Anima commercially?
- Not without separate permission. Anima is released under the CircleStone Non-Commercial, which allows research and personal use only.
- Is Anima open source?
- Anima is open-weight, not open source in the strict sense. You can download the model files, but the CircleStone Non-Commercial adds conditions that the Open Source Initiative definition does not allow.
- Can I run Anima on my own computer?
- Yes. The smallest variant, Anima-Turbo v1.1 (BF16), needs about 6.44 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Anima released?
- Anima was released by CircleStone Labs on May 14, 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Anima model card (Hugging Face) huggingface.co
- CircleStone Labs Non-Commercial License (Hugging Face) huggingface.co
- Anima repository file list and history (Hugging Face) huggingface.co
- NVIDIA Open Model License Agreement nvidia.com
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Z-Image Tongyi MAI | 6.15B | – | Apache 2.0 (Permissive) | 16 GB Laptop or phone | |
| MOSS-Image OpenMOSS | 5.08B | – | Apache 2.0 (Permissive) | 4.2 GB Laptop or phone | |
| Qwen-Image-2.1 Alibaba Qwen | 7.1B | – | Qwen Research License (Non-commercial) | 20 GB Gaming GPU | |
| FLUX 3 Image Klein Black Forest Labs | 5.6B | – | FLUX Kommunity (Conditional) | 8 GB Laptop or phone |