Today marks the official release of Wan2.5-Preview, a new AI model with a revolutionary architecture and powerful features designed to reshape the future of visual generation. The new model achieves significant breakthroughs in multimodal processing, video generation, and image editing.
Native Multimodal Architecture and Deep Alignment
Wan2.5-Preview utilizes a new unified framework for understanding and generation, allowing for flexible input and output of text, images, video, and audio. By jointly training on this multimodal data, the model achieves stronger modal alignment, which is crucial for synchronized audio-visual output and precise instruction following. Additionally, the model is optimized through Reinforcement Learning from Human Feedback (RLHF) to ensure that the generated images and video dynamics align with human aesthetic preferences.
Video Capabilities: Synchronized A/V and Cinematic Aesthetics
Wan2.5-Preview introduces several innovations in video generation:
- Synchronized A/V Generation: It natively supports high-fidelity, highly consistent video generation, while simultaneously generating audio, including multiple voices, sound effects, and background music (BGM).
- Controllable Multimodal Input: Users can use text, images, and audio as input sources to achieve endless creative combinations.
- Cinematic Aesthetics: The model generates 10-second 1080p HD videos with strong dynamics and structural stability. It also features an upgraded cinematic control system for creating works with a true film aesthetic.
Image Capabilities: Creativity and Precise Control
Wan2.5-Preview also brings significant improvements to image generation and editing:
- Advanced Image Generation: The model has a much-improved ability to follow instructions, capable of generating realistic images, diverse artistic styles, creative layouts, and professional diagrams.
- Image Editing: It supports conversational and instruction-based image editing with pixel-level precision for complex tasks such as multi-concept blending, material transformation, and product color swapping.
The release of Wan2.5-Preview signals a new phase in AI visual generation technology. Its powerful multimodal capabilities and precise control features will provide developers and creators with unprecedented tools.
© Copyright notes
The copyright of the article belongs to the author, please do not reprint without permission.
