| --- |
| base_model: |
| - circlestone-labs/Anima |
| license: other |
| license_name: circlestone-labs-non-commercial-license |
| license_link: LICENSE.md |
| tags: |
| - text-to-image |
| - anime |
| - webgpu |
| - quantized |
| --- |
| |
| # Anima for the browser |
|
|
| Weights for **Anima Studio**, which runs [Anima](https://huggingface.co/circlestone-labs/Anima) |
| (by CircleStone Labs and Comfy Org) entirely in a web browser with WebGPU. The page downloads |
| these files once, keeps them in the browser's storage, and generates images on the visitor's own |
| graphics card. Nothing is sent to a server. |
|
|
| **Unofficial.** These are modified (repackaged and quantized) versions of Anima, not an official |
| CircleStone Labs release. Non-commercial use only: see [LICENSE.md](LICENSE.md) and [NOTICE](NOTICE). |
|
|
| ## Files |
|
|
| Image models (2B), one file per version and size: |
|
|
| | Version | Family | Compressed (int8) | Extra compressed (w4a8) | |
| | --- | --- | --- | --- | |
| | Anima Turbo v1.1 | Turbo | 2.42 GB | 1.71 GB | |
| | Anima Turbo v1.0 | Turbo | 2.42 GB | 1.71 GB | |
| | Anima Aesthetic v1.1 | Aesthetic | 2.42 GB | 1.71 GB | |
| | Anima Aesthetic v1.0 | Aesthetic | 2.42 GB | 1.71 GB | |
| | Anima Aesthetic v1.0b | Aesthetic | 2.42 GB | 1.71 GB | |
| | Anima Base v1.0 | Base | 2.42 GB | 1.71 GB | |
|
|
| Shared by every version: |
|
|
| | Part | Compressed (int8) | Extra compressed (w4a8) | |
| | --- | --- | --- | |
| | Text encoder (Qwen3 0.6B) | 0.61 GB | 0.44 GB | |
| | Image decoder (Qwen-Image VAE, decoder only) | 0.05 GB | | |
|
|
| `manifest.json` lists the files, their sizes and each version's recommended settings: Turbo runs |
| at CFG 1 with 8 steps (ER-SDE, beta schedule); Aesthetic and Base use guidance (CFG 4 / 4.5) with a |
| negative prompt and 30 steps (ER-SDE, simple schedule). `tokenizers/` holds the Qwen and T5 |
| tokenizers. |
|
|
| - **int8**: ComfyUI `int8_tensorwise` with ConvRot (256-group Hadamard rotation), as in |
| Comfy-Org's MiniMax-H3 int8_convrot release. |
| - **w4a8**: ComfyUI `asym_w4a8_int8` ("mixed": first and last image-model blocks stay int8), as in |
| Kijai's MiniMax-H3 w4a8_mixed release. |
| - The decoder's causal 3D convolutions are collapsed to 2D for single images. The text encoder's |
| final layer stays bf16. |
|
|
| ## License |
|
|
| The CircleStone Model is licensed by CircleStone Labs LLC under the CircleStone Non-Commercial |
| License. Copyright CircleStone Labs LLC. See [NOTICE](NOTICE) for the full attribution notice, |
| the NVIDIA Open Model License (Anima is a Derivative Model of Cosmos-Predict2-2B-Text2Image), and |
| the Apache 2.0 Qwen components. |
|
|
