On February 18, Kunlun Wanwei announced the release of SkyReels-V1, China's first open-source video generation model for AI short drama creation, and SkyReels-A1, the country's first SOTA-level expression and motion controllable algorithm based on a video foundation model. This initiative aims to address current issues with AI video generation models, such as lack of openness, high costs, and difficulty of use, promoting innovation and development in the AI short drama industry.

SkyReels-V1 is China's first open-source video generation model designed for AI short drama creation. This model achieves fine control over the details of character performances by integrating multiple generation modules, including lip-sync generation, expression generation, and body movement generation. SkyReels-V1 is trained and fine-tuned using high-quality Hollywood data at a scale of tens of millions, significantly enhancing the controllability of expression and body movement generation.

SkyReels-V1 supports 33 intricate character expressions and over 400 natural action combinations, enabling a high degree of realism in emotional expression. The generated videos possess cinematic lighting aesthetics and image quality, capable of producing film-level micro-expression performances. Additionally, SkyReels-V1 supports both text-to-video and image-to-video generation, making it the model with the largest parameters for image-to-video support among open-source video generation models.

Technically, SkyReels-V1 relies on Kunlun Wanwei's self-developed high-quality data cleaning and manual annotation pipeline, building a dataset of tens of millions of high-quality films, TV shows, and documentaries. The team also developed a "Human-Centric" multimodal large model for video understanding, enhancing the understanding of characters in videos.

SkyReels-A1 is the first SOTA-level algorithm for controllable expressions and motions based on a video foundation model. This algorithm supports video-driven, film-level expression capture, achieving high-fidelity micro-expression restoration. SkyReels-A1 can generate highly realistic dynamic videos of characters based on any body proportion, with its realism stemming from deep restoration of multi-dimensional details such as changes in character expressions, emotions, skin texture, and body movements.

SkyReels-A1 supports expression control generation for profiles, enabling more realistic eyebrow and eye micro-expressions, as well as greater head and natural body movements. Compared to Runway's Act-One, SkyReels-A1 maintains the integrity of characters in the generated videos, with more realistic performance details, allowing for a natural and perfect fusion of expressions and body movements.

Kunlun Wanwei has consistently advocated for open-source practices, promoting technological equity. Since August 2023, the company has successively open-sourced several models, including the large language model "Skywork" Skywork-13B series, the full-process R&D toolkit for digital agents AgentStudio, and the "Skywork Large Model 3.0," a 400 billion parameter MoE super model. The open-sourcing of SkyReels-V1 and SkyReels-A1 marks a first in the AI short drama industry, aiming to provide users with lower-cost and more controllable AIGC capabilities.

Kunlun Wanwei's Chairman and CEO, Fang Han, stated that the emergence of AIGC capabilities will significantly reduce film production costs and promote cultural equity. In the future, Kunlun Wanwei will continue to open-source more video generation models and algorithms, driving the continuous development and prosperity of the AI short drama ecosystem.

Open-source addresses:

SkyReels-V1

https://github.com/SkyworkAI/SkyReels-V1

SkyReels-A1

https://github.com/SkyworkAI/SkyReels-A1

Technical Report:https://skyworkai.github.io/skyreels-a1.github.io/report.pdf

Official SkyReels Address:skyreels.ai