--- base_model: - circlestone-labs/Anima license: other license_name: circlestone-labs-non-commercial-license license_link: LICENSE.md tags: - diffusion-single-file - comfyui - anima - dc-gen --- ![Battera-Anima-DC sample](https://huggingface.co/FA770/Battera-Anima-DC/resolve/main/images/sample1.webp) # Battera-Anima-DC v1.0 **Battera** is a Deep-Compressed [Anima](https://huggingface.co/circlestone-labs/Anima): the Anima diffusion transformer, moved onto a custom 32× compressed autoencoder ([Battera-VAE-f32c128](https://huggingface.co/FA770/Battera-VAE-f32c128) ) with the [DC-Gen](https://arxiv.org/abs/2509.25180) recipe. The latent grid is 4× smaller than stock Anima (1024 tokens instead of 4096 at 1024×1024), so generation is roughly **4× faster** with the same text encoder and the same prompting style. > [!IMPORTANT] > Battera is **not** a drop-in Anima checkpoint. It needs: > - its own VAE, **[Battera-VAE-f32c128-v1.0](https://huggingface.co/FA770/Battera-VAE-f32c128)** (the Qwen-Image VAE will not work) > - a custom node (ComfyUI) or extension (Forge Neo), linked below > > The transformer and the VAE are a matched pair. Always use the same version number for both. This is an unofficial derivative model. It is not affiliated with or endorsed by CircleStone Labs. ## Files | File | Size | Put it in | |---|---|---| | `Battera-Anima-DC-v1.0.safetensors` | 4.18 GB | ComfyUI: `models/diffusion_models` / Forge Neo: `models/Stable-diffusion` | | `Battera-VAE-f32c128-v1.0.safetensors` (from the [VAE repository](https://huggingface.co/FA770/Battera-VAE-f32c128) ) | 1.75 GB | `models/vae` (ComfyUI) / `models/VAE` (Forge Neo) | Text encoder: the same **Qwen3 0.6B** used by stock Anima (`models/text_encoders` in ComfyUI, `models/text_encoder` in Forge Neo). ## Usage ### ComfyUI Install [ComfyUI-Anima-DC](https://github.com/FA7770/ComfyUI-Anima-DC) into `custom_nodes`, then: | Node | Input | Connect from | |---|---|---| | Anima-DC Empty Latent | model | Anima-DC Loader | | KSampler | model | Anima-DC Loader | | KSampler | latent_image | Anima-DC Empty Latent (or Anima-DC VAE Encode for img2img) | | KSampler | positive / negative | CLIP Text Encode (Load CLIP → Qwen3 0.6B) | | Anima-DC VAE Decode | samples / vae | KSampler / Anima-DC VAE Loader | The loader uses Anima's default shift (3.0). To change it, put **Anima-DC Flow Shift** between the loader and KSampler. **Example workflow**: download `images/sample_ComfyUI.png` from this repository and drag and drop it onto ComfyUI to load the full workflow. ### Forge Neo Install [sd-webui-anima-dc](https://github.com/FA7770/sd-webui-anima-dc) from **Extensions → Install from URL** and restart. Then: 1. Checkpoint: `Battera-Anima-DC-v1.0` 2. VAE / Text Encoder: `Battera-VAE-f32c128-v1.0` **and** Qwen3 0.6B 3. Leave **Shift** at the Anima default (3) The console shows `[Anima-DC] loader hooks installed` at startup when the extension is active. **Example settings**: download `images/sample_forge-neo.png` from this repository, load it in the **PNG Info** tab and send it to txt2img. The prompt and sampler settings are restored; select the checkpoint and the VAE / Text Encoder manually. ## Recommended settings | Setting | Value | |---|---| | Resolution | multiples of 32, around 1024² (other sizes are rounded down) | | Sampler / scheduler | Euler / Simple | | Steps | 24–30 | | CFG | 4-6 | | Shift | 3 (Anima default) | Prompting is the same as stock Anima (Danbooru-style tags and/or natural language, `@artist`, quality and score tags). See the [Anima model card](https://huggingface.co/circlestone-labs/Anima) for details. ## Speed RTX 4090, 1024×1024, batch 4, identical sampler settings, ComfyUI (total prompt time): | Model | Time | |---|---| | Anima (stock) | 33.6 s | | Battera-Anima-DC v1.0 | 8.4 s | ## Limitations - Fine detail is somewhat lower than stock Anima. The adaptation is not perfect, so the same prompt and seed do not give the same image as Anima; composition can change. - Stock Anima LoRAs work, but not perfectly. LoRAs retrained on Battera are recommended. - Do not quantize the VAE to FP4; the output breaks. bf16 and fp32 work. - ControlNet and Anima's reference feature are not supported. ## License The model weights are a Derivative of the CircleStone Model and are distributed under the **CircleStone Labs Non-Commercial License v1.2** (`LICENSE.md`). Non-commercial use only; all restrictions of the license apply to this model. Anima is built on NVIDIA Cosmos; see the Anima model card for the NVIDIA Open Model License terms that apply to derivatives. Attribution notice (also in `NOTICE.txt`): > The CircleStone Model is licensed by CircleStone Labs LLC under the CircleStone Non-Commercial License. > Copyright CircleStone Labs LLC. > IN NO EVENT SHALL CIRCLESTONE LABS LLC BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION > OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL. > > This model has been modified by FA770 from the original CircleStone Model (Anima). This license covers the transformer only. Battera-VAE-f32c128 is a separate model with its own license (Apache License 2.0), and the ComfyUI node and the Forge Neo extension are separate projects under the Apache License 2.0. ## Acknowledgements - [Anima](https://huggingface.co/circlestone-labs/Anima) by CircleStone Labs - [DC-Gen](https://arxiv.org/abs/2509.25180) by MIT HAN Lab / NVIDIA - [kohya-ss/sd-scripts](https://github.com/kohya-ss/sd-scripts), used for training