What is Anima?

Anima is a 2B parameter image generation model created through a collaboration between CircleStone Labs and Comfy Org.

It is based on NVIDIA Cosmos, and is a pure anime model trained on several million anime images and art images.

It is one of the models drawing a lot of attention as a possible migration target from SDXL-era anime models.

The license is CircleStone Labs Non-Commercial License.
The Anima model itself, derivative models, and LoRAs are non-commercial. Generated images can be used commercially.

Open this article's workflows on RunpodRecommended : RTX A5000RTX 3090RTX 4090How it works

Model Download

anima-base-v1.0 is the untuned base model. anima-aesthetic-v1.1 is fine-tuned on high-quality images. If you just want to try generating images, this is generally the one to use.

📂ComfyUI/
└── 📂models/
    ├── 📂diffusion_models/
    │   ├── anima-base-v1.0.safetensors
    │   └── anima-aesthetic-v1.1.safetensors
    ├── 📂text_encoders/
    │   └── qwen_3_06b_base.safetensors
    └── 📂vae/
        └── qwen_image_vae.safetensors

Prompt

Anima is trained on Danbooru-style tags, natural language, and combinations of both.

Because the text encoder is smart, you do not need to be as strict as with Pony Diffusion V6, but it may be useful to put quality tags like the following at the beginning of the prompt.

masterpiece, best quality, score_9, safe,

text2image

anima_aesthetic_v1_1.json
  • Recommended resolution is 512px to 1536px.
  • Samplers with a little variation, such as er_sde and euler_ancestral, are recommended.

Anima LLLite

ControlNet-LLLite is a lightweight ControlNet developed by kohya.

Model Download

Choose a suitable model for the control image you want to use.

Models marked v2 were trained for Anima-Base v1.0. The others were trained for the older Anima Preview3, so they may be a little less effective.

📂ComfyUI/
└── 📂models/
    └── 📂model_patches/
        └── anima-lllite-*.safetensors

anima-lllite-any-test-like-v2

anima_lllite_any_test_like_v2.json
  • anima-lllite-any-test-like-v2 combines several types of control in a single model. It can use rough sketches or line art to guide the composition, or colorize grayscale images.
  • In this workflow, Canny extracts the outlines from the input image and uses them as the control image.