Version timeline
- AniSora V1: built on CogVideoX-5B, with full training and inference code and cost-effective deployment on an RTX 4090.
- AniSora V2: built on Wan2.1-14B; weights released under Apache-2.0 on 11 July 2025, with native Huawei Ascend 910B NPU support.
- AniSora V3: weights released under Apache-2.0 on 27 August 2025, adding arbitrary-frame inference, character 3D, style transfer, and multimodal guidance.
- AniSora V3.1: released 4 September 2025 with a larger motion range.
- AniSora V3.2: released 23 September 2025, trained on Wan2.2 with 8-step inference.
- AniSora AnyMask: released 31 October 2025 for image-to-video with temporal and spatial masks.
Research behind AniSora
The project also releases a data-cleaning pipeline and an anime benchmark of 948 clips, labeled by action with 10 to 30 clips per label.
- AniSora: Exploring the Frontiers of Animation Video Generation in the Sora Era (arXiv 2412.10255), accepted at IJCAI 2025.
- Aligning Anime Video Generation with Human Feedback (arXiv 2504.10044), the first RLHF framework for anime video generation.
- AniMe (arXiv 2508.18781), agent research for long animation production.
Which AniSora version should you use?
Setup guides: see the AniSora V3 and AniSora V3.2 pages on this site, which follow the official READMEs.
- Best quality and speed on large GPUs: V3.2.
- Limited VRAM: the 12 GB build of V3.1.
- Ascend NPUs: V2.
- Masked or localized animation: AnyMask.
Spelling variants
People also search for Index AniSora as AniSora AI, Ani Sora, or Anizora. They refer to the same Bilibili project.
Try anime image-to-video without a GPU
The hosted generator on AniSora Studio currently animates your image with Kling 2.1 image-to-video through fal. It does not run the open-source AniSora weights, and the model used is always disclosed before you spend credits.
AniSora Studio is an independent product. The official Index-anisora GitHub repository states that it is the only official homepage of the AniSora project.