Index AniSora is Bilibili's open source animation video generation project. It is systematically constructed around animation style video generation, spatiotemporal control, character consistency and specialized evaluation benchmarks. The goal is not general realistic video, but animation content production.

Index AniSora

Core parameters and statistics

[A brief comment in one sentence]: It is not an animation skin for a general video generation model, but a complete set of data, models and evaluation systems specially established by Bilibili for animation video production.

[Publicity Verification]: The official warehouse directly writes itself as “the most powerful open-source animated video generation model”. Of course, this kind of wording has a promotional component, but at least the direction is not wrong, because the project, from papers, data pipelines, evaluation benchmarks to multi-copyrights, is indeed centered around "animated video" rather than pan-video generation. The pain points it hits are clear: realistic video models often move erratically in anime style, physical logic and artistic exaggeration conflict, and character consistency is poor.

Projects	Public Information
Official positioning	Open source animation video generation model and system
Current Mainline	V3.2
Paper	IJCAI 2025 Accepted
Open source portal	GitHub, Hugging Face, ModelScope
Data scale	Papers and model cards publicly mention 10 million levels of high-quality animation data
Benchmark	AniSora-Benchmark, including 948 animated video clips
Community size	GitHub about 2.5k stars, 147 forks

Expert’s point of view: The most valuable thing about this project is not the single model checkpoint, but the “construction of an evaluation and data cleaning system specifically for animation scenes”. Because of the style, exaggerated movements and character modeling stability of animated videos, it is not suitable to directly apply the benchmark of the real world.

User and market recognition

AniSora's current recognition comes more from the open source community and animation content technology circles. GitHub has about 2.5k stars and Hugging Face has about 227 likes. It is not a national hit, but for a vertical video generation project, it has entered the stage of "some people continue to pay attention and some are preparing for secondary development".

User and market recognition: The official adaptation scenarios are very clear, including animated series, nationally created animations, comic-adapted VTuber content, animated PVs, and ghost and animal works. This shows that it is not aimed at all video teams, but is clearly betting on the animation content production chain.

Publicity Verification: If you want to generate realistic commercials or brand live-action videos, AniSora does not have an advantage; but if you want animation character movements, animation lens control and style transfer, its product positioning is more focused than the general video model.

Hidden benefits: What the animation team fears most is the high cost of early verification and the need for a lot of manual testing of character movements. The advantage of AniSora is that it makes it cheaper to publish a version of animated shots first.

Cost advantage

Free truth: The fact that the project weight and code are open source does not mean that everyone can use it at low cost. Public video generation, especially 3D characters, style transfer and long shot control, requires high memory, inference time and storage costs. The so-called "open source and free" mainly saves license fees, not training and inference costs.

Cost layer	Disclosure	What it actually means
C-side/individual	No finished service for ordinary consumers	More like a prototype project than an out-of-the-box product
Developer/API	Weights are open source and can be downloaded	You must bear the cost of model deployment, graphics card, inference and parameter adjustment by yourself
Enterprise	No public commercial package	If you want to enter production, you need to build your own platform and content review process

Hidden Cost: Animated video generation especially requires "style consistency" and "controllability". The license fees you save are often spent on graphics cards, inference queues, parameter exploration, and manual screening.

Hidden benefits: For animation studios or virtual content teams, the most valuable thing is that they can verify actions and shots faster without having to go through the complete art and storyboarding every time.

Main functions

• Animation Video Generation: For animation style rather than realistic videos.
• Arbitrary frame reasoning and timing control: more suitable for animation lens connection and intermediate frame construction.
• Character 3D video generation: Generate multi-angle dynamic videos from frontal character images to help keep characters stable.
• Video Style Migration: Migrate the original video to the target animation style.
• Multi-modal guidance: Supports pose, depth, line art, audio and other guidance methods to control actions.
• Ultra-low resolution video super-resolution: Reduce the sampling cost, and then increase the picture to a higher resolution.

Expert View: The biggest hidden linkage between these functions is that it puts both "character consistency" and "action control" into the same production chain. The most common problem with universal video models is that the characters' faces change when they move, and animation content is most afraid of this.

Model and version evolution

The rhythm of AniSora's version is very clear, and almost every stage is solving a hard problem in animation production.

Public version main line

V3: 2025-08-27 The weights are disclosed, and arbitrary frame reasoning, character 3D video, style transfer and multi-modal guidance are added, which is a key node from "able to generate" to "able to control".

V3.1: 2025-09-04 Focus on strengthening range of motion control, indicating that the team has begun to deal with more detailed dynamic naturalness issues.

V3.2: 2025-09-23 Based on stronger base model training, the number of inference steps is reduced to 8, which means that it is chasing speed and effect at the same time.

Current Limitations: Although the version line is very clear, there is no stable commercial support and SLA of traditional SaaS. Production bounded access is more suitable for fixing specific weight versions, rather than chasing the latest all the way.

Technical advantages

Index AniSora is a cross between [basic large model/API infrastructure] and vertical video production tools.

Performance and Throughput: The public page does not give TTFT, RPM or cloud concurrency limits; but the official clearly states that V3 can generate 5 seconds of 360p animation video footage in 8 seconds, which is one of the few public information that can directly reflect the inference efficiency. Real throughput for high resolution and complex control scenarios still depends on the deployment context.

Mechanics -> Effects -> Scenes:

Data pipeline specifically for animation: Construct high-quality training data through scene detection, optical flow, aesthetic scoring, etc. The effect is that the model understands the animation language better, and is most suitable for VTuber, PV, comic modification and second-generation content.

Space-time mask and multi-modal control: The effect is to enable more detailed control of actions and local areas, rather than random generation of the entire segment, suitable for animation lens design and specific action demonstrations.

Specialized Benchmarks Aligned with Human Preferences: The value of AniSora-Benchmark is that it evaluates not a general realistic world, but animation aesthetics and action consistency.

Adaptation Boundary: It is best at animation and two-dimensional content; it is least good at realistic commercial images and serious videos that require real mouth shapes and real photography grammar.

How to use

git clone https://github.com/bilibili/Index-anisora.git
cd Index-anisora
# Select the corresponding directory according to the target version, such as anisoraV3.2 or anisoraV3

Note: The official warehouse and Hugging Face have disclosed multiple version directories and weight entries. The actual deployment depends on different versions. The specific commands are subject to the corresponding subdirectory README.

Entrance method:

Entrance	Applicable objects	Description
GitHub training/inference code	R&D team	Suitable for running training and inference by yourself
Hugging Face weights	Model engineering team	Suitable for downloading and experimentation
ModelScope	Chinese developer ecosystem	Suitable for domestic downloads and secondary integration

Boundary of human-machine collaboration: Shot drafts and motion exploration can be automated, but before they are actually used for commercial animation or character IP content, the films must still be reviewed manually, focusing on character stability, style consistency, and copyright risks.

Product Pricing

There is currently no publicly available commercial subscription form. Its cost model is not a "monthly fee", but a typical open source video model cost: graphics card, storage, inference time, manual screening and content review.

C-side/Personal: There is no ready-made lightweight consumer version.

Developer/API: The code and weights are open source, but the calling cost comes from computing power.

Enterprise: If you plan to build an internal animation asset platform, you must also include queue management, permissions, review, and material copyright management in your budget.

The truth about free: The more complex the animation video control is, the less likely it is to solve the total cost problem by just relying on the words "open source and free".

Application scenarios

• VTuber and virtual character content: High demand for character stability and motion control.
• Anime PV and promotional video: Quickly test shots, test momentum, and test style.
• Comic Adaptation and Secondary Animation Prototypes: Suitable for dynamic drafts and concept demonstrations first.
• Animation Technology Research and Evaluation: The project comes with a special benchmark to facilitate vertical comparison.

Dimensionality reduction attack scenario: When the target is animation instead of pan-video, its verticality will obviously exceed the general Vincent video model.

Current Limitations: It is not optimal for realistic videos, long continuous narratives, and strong lip-synced dialogue.

Applicable people

• Animation Studio and Virtual Content Team: Want to speed up shot verification and motion exploration.
• Video Generation Research Team: Requires public weights of animation categories and dedicated benchmarks.
• Independent creators and secondary developers: Hope to try animation video generation in the open source system.

Persuasion Scenario:

• People who just want to quickly produce ready-made marketing short videos: It focuses on R&D and prototype projects.
• People without graphics card resources: The deployment threshold is not low.
• People who make realistic videos of real people: Its world view is not optimized for realistic videos.

Summary and Outlook

The real value of Index AniSora is not as simple as "proving that animation can also produce videos", but that it turns animation video generation into a systematic project with data pipelines, version main lines, and dedicated evaluation. For vertical content production, this kind of specialized construction is more meaningful than minor repairs to the general model.

[Procurement/Adoption Risk Assessment]: If the team is doing animation VTuber, PV, secondary creation or character video, AniSora is worthy of continued attention; if the business goal is realistic advertising and pan-video generation, the direction of capabilities may be mismatched. Before adoption, focus on verifying graphics card resources, target style suitability, character stability, and secondary copyright boundaries. Don’t let the word “open source” cover up the true implementation costs.

Related tools: runway, pika

Version Info

• AniSora V3.2 ：The latest weight mainline disclosed in the official update description is based on stronger Wan2.2 training and compresses the number of inference steps to 8 steps, continuing to retain the functions of arbitrary frame inference and character 3D video generation. （2025-09-23）
• AniSora V3.1 ：Improves the range of motion capabilities and recommends that the motion score be set to 2.0 to 4.0, which is an important enhanced version of animation momentum control. （2025-09-04）
• AniSora V3 ：The weights are exposed in Hugging Face and ModelScope, and new arbitrary frame inference, character 3D video generation, video style transfer, multi-modal guidance and ultra-low resolution video super-resolution are added. （2025-08-27）