Alibaba has unveiled Wan2.5-Preview, a new video generation model that integrates audio creation capabilities. This advanced system can produce videos with synchronized sound, seamlessly combining various media types. The launch positions Alibaba as a strong competitor in the rapidly evolving field of AI-powered video synthesis, directly challenging Google’s recent advancements with Veo 3.
Key Takeaways
- Alibaba’s Wan2.5-Preview generates videos with synchronized audio, including voices, sound effects, and music.
- The model supports a multimodal architecture, processing text, images, video, and audio within a unified system.
- Wan2.5-Preview offers 1080p resolution for videos up to 10 seconds long.
- The platform wan.video allows users to create and edit images via voice commands, similar to OpenAI’s Sora interface.
- Unlike its predecessor, Wan2.5-Preview is not open-source, with usage available through a subscription or credit system.
Advanced Multimodal Capabilities
Wan2.5-Preview is built on a multimodal architecture, enabling it to process text, images, video, and audio within a single, cohesive system. Alibaba states that training the model on all data types simultaneously leads to improved synchronization and coherence across different media. While details about the architecture are sparse, Alibaba mentions the use of reinforcement learning with human feedback, describing Wan2.5-Preview as a significant step towards a ‚world model‘.
Synchronized Audio and Video Generation
The model’s ability to create videos with synchronized audio is a key feature. This includes generating multiple voices, sound effects, and background music to accompany the visuals. The generated videos achieve a 1080p resolution and can be up to 10 seconds in length. Demonstrations show the model’s capacity to combine various clips, though initial observations suggest minor synchronization issues between audio and visual elements, particularly with musical timing. The model also faces challenges in maintaining consistent facial features across generated video sequences.
User Interaction and Image Editing
Users can input text, images, or audio to generate videos. For instance, a user can upload a photo and use a text prompt to create a video with matching music, with Alibaba promising "cinematic aesthetics" and a "cinematic control system." Beyond video generation, Wan2.5-Preview, accessible via wan.video, also allows for image creation and editing. The platform can produce photorealistic images, various artistic styles, and diagrams. Image editing can be controlled through voice commands, enabling tasks like changing product colors or merging different concepts.
Competitive Positioning and Accessibility
With its integrated audio generation, Wan2.5-Preview is positioned to compete with models like Google’s Veo 3, which was the first video model capable of generating accompanying audio. Unlike previous Alibaba models, Wan2.5-Preview is not available as open-source, a decision that has not been met with detailed explanation from Alibaba. Access to Wan2.5-Preview is provided through the wan.video platform, with monthly subscriptions starting at $6.50 or via credit purchases, making individual clips cost between 13 and 25 cents. API usage is priced between 5 and 15 US cents per second of video, significantly lower than Veo 3’s API costs.
KI Snack
