On May 19, 2025, Bilibili (B站) announced the open-sourcing of its latest animated video generation model—Index-AniSora. This innovative technology has brought a revolutionary breakthrough to the generation of secondary-style videos. Index-AniSora supports one-click generation for various styles of secondary-style video clips, including anime series, domestic creations, comic adaptations, VTubers, animation PVs, and meme animations, greatly enhancing the production efficiency and quality of animated content.
The technical principle of Index-AniSora is based on Bilibili's AniSora model, which has been accepted by the International Joint Conference on Artificial Intelligence (IJCAI) in 2025. On this basis, Bilibili further proposed the first reinforcement learning framework specifically designed for secondary-style video generation. By aligning the optimization of animation video generation through human feedback, it significantly improves the overall quality of the generated content.
In terms of technical implementation, Bilibili's research team constructed the first high-quality reward dataset for the anime domain, containing 30,000 manually labeled anime video samples. This dataset evaluates video quality from two aspects: visual appearance and visual consistency, covering multiple dimensions such as visual smoothness, visual motion, visual appeal, text-video consistency, image-video consistency, and character consistency. Based on these dimensions, the research team proposed AnimeReward, a multi-dimensional high-confidence reward system specifically designed for aligning anime video generation.
To further enhance the alignment performance of the model, the research team proposed Gap-Aware Preference Optimization (GAPO), incorporating the preference gap between positive and negative sample pairs into the loss function to improve the efficiency and final performance of the alignment training. Experimental results show that the model optimized by AnimeReward and GAPO significantly outperforms baseline models and supervised fine-tuning (SFT) models in multiple evaluation dimensions, generating animation videos that are more in line with human preferences.
Bilibili's open-source project not only brings new technological breakthroughs to the field of animated video generation but also provides valuable resources and tools for developers and enthusiasts. Through Index-AniSora, users can easily convert their favorite comics into vivid animation effects, supporting various niche art styles, with richer results, completely bidding farewell to the "PPT animation" era. The openness of this technology will undoubtedly promote further development in secondary-style content creation, bringing more possibilities for anime enthusiasts and creators.
Address:
https://github.com/bilibili/Index-anisora/tree/main
Model link:
https://modelscope.cn/models/bilibili-index/Index-anisora
Experience link:
https://modelscope.cn/studios/bilibili-index/Anisora
