量子位·2025-05-15 06:26Core Viewpoint - Alibaba has launched an open-source video generation and editing model called Wan2.1-VACE, which claims to be the most comprehensive in the industry for consumer-grade graphics cards [1][2]. Group 1: Model Features - Wan2.1-VACE offers a wide range of functionalities, described as "All in one," allowing users to experience various video generation capabilities within a single model [2]. - The model is available in two versions: a 1.3B version that supports 480p resolution and a 14B version that supports both 480p and 720p resolutions [7]. - The model supports multiple video generation methods, including text-to-video, image-to-video, and video-to-video [8]. Group 2: Video Generation Capabilities - The model demonstrates smooth video generation based on reference images, showcasing natural movements and harmonious composition [11]. - Users can create videos with specific prompts, such as a scene featuring a girl playing with a cartoon snake, highlighting the model's ability to capture detailed atmospheres [12]. Group 3: Editing Features - Wan2.1-VACE includes essential editing functionalities, addressing the common issue where most AI video generators do not achieve 100% success on the first attempt [16]. - The model allows for "creation from nothing," enabling users to generate scenes and then edit them by adding or modifying elements [17][21]. - It supports advanced editing features such as pose transfer, motion control, and scene extension [22]. Group 4: User Experience and Performance - Users have begun testing the model, with positive feedback on its ability to change video aspect ratios and maintain detail control during pose and facial expression transfers [27][29]. - The computational efficiency of Wan2.1-VACE varies across different GPUs, with specific performance metrics provided for various models and resolutions [26]. Group 5: Community Engagement - The article encourages community interaction, inviting readers to express interest in further testing and sharing their experiences with the model [32].