Index-AniSora is an open-source framework for generating anime-style videos, targeting anime creators, researchers, and enthusiasts. It offers one-click video creation across various anime styles and provides tools for dataset expansion and specialized evaluation, aiming to advance the state-of-the-art in animation video generation.

How It Works

AniSora leverages the CogVideoX-5B and enhanced Wan2.1-14B foundation models. It incorporates a spatiotemporal mask module for precise video control, enabling image-to-video generation, frame interpolation, and localized image-guided animation. The system also includes an end-to-end data pipeline for training data expansion and a specialized benchmark system with tailored evaluation models and algorithms for animation generation.

Quick Start & Requirements

• Installation: Code is available in anisoraV1_infer, anisoraV2_gpu, anisoraV2_npu, anisora_rl, data_pipeline, and reward directories.
• Prerequisites: AniSora V2 is optimized for Huawei Ascend 910B NPUs, with V1 deployable on RTX 4090. Specific CUDA versions and Python versions are not explicitly stated but are implied by the underlying models.
• Resources: V1 is described as cost-effective on an RTX 4090. V2 is trained entirely on domestic chips (Huawei Ascend 910B).
• Links: GitHub, Hugging Face, Model Scope, Paper, Video Gallery

Highlighted Details

• AniSora V2 weights are licensed under Apache 2.0.
• Features RLHF framework for anime video generation (AniSoraV1.0_RL).
• Includes an anime-optimized benchmark system with a dataset of 948 animation video clips.
• Supports localized region guidance and temporal control (first/last frame guidance, keyframe interpolation).

Maintenance & Community

• Project is actively updated with new versions (V1, V2, V3 preview).
• Paper accepted by IJCAI'25.
• Community engagement is encouraged via group chats.

Licensing & Compatibility

• AniSora V2 weights are licensed under Apache 2.0.
• No explicit mention of licensing for other components or older versions.

Limitations & Caveats

• Access to benchmark data requires agreeing to Bilibili's terms and submitting a PDF with a handwritten signature from an affiliated institution.
• The project is presented as Bilibili's "gift to the anime world," with a strong emphasis on their official homepage.