In July 2025, Bilibili rolled out a significant update to its open – sourceanime video generation model, unveiling AniSora V3. This release is partof the Index – AniSora project, aiming to empower creators in anime, manga,and VTuber fields.
Technical Upgrades: AniSora V3 is built on Bilibili’s prior CogVideoX – 5Band Wan2.1 – 14B models, integrating the Reinforcement Learning with HumanFeedback (RLHF) framework. This leads to remarkable improvements in visualquality and motion consistency of generated videos. It can generate variousanime video scenes with a single click.
The core upgrades are multi – Faceted. First, the optimization of theSpatiotemporal Mask Module enhances spatiotemporal control, handling complexanimation tasks. Second, the dataset is expanded with over 10 million high -quality anime video clips, and a new data cleaning Pipeline is added. Third,hardware optimization includes native support for Huawei Ascend910B NPU,trained on domestic chips and boosting inference speed by 20%. Fourth, multi -task learning capabilities are enhanced, making it ideal for manga adaptationsand VTuber content creation.
In VBench and double – blind subjective tests, AniSora V3 reached topindustry lEVEls in role consistency and motion smoothness, especiallyexcelling in complex actions.
Open Source Ecosystem: On July 2, 2025, the complete training andinference code of AniSora V3 was updated on GitHub. Developers can accessmodel weights and a dataset via Hugging Face. The introduction of the firstRLHF framework for anime video generation ensures outputs match humanaesthetics. CommUnity developers are creating custom plugins based on V3.
Application Scenarios: AniSora V3 supports a wide range of anime styles,covering 90% of application scenarios. It includes single image to videogeneration, manga adaptation, VTuber and game applications, and high -resolution output. AIbase testing shows it reduces artifact issues andshortens generation time.
AniSora V3 reduces the barrier to anime creation, catering to independentcreators and small teams. It fills a market gap in the anime field compared togeneral video generation models, and presents a different technical approachcompared to ByteDance’s EX – 4D.
