wan3.0-video

Wan3.0-Video is the flagship all-in-one video generation model of the Wanxiang 3.0 series. It adopts an All-in-One design and can cover all task types such as Vincent videos, Tu videos (first frame/first and last frame), and Reference videos without switching models. Its core breakthrough lies in the ability to understand multi-modal materials: it supports up to 20 multi-modal material references at a time, covering pictures, videos, audios, PDF/PPT/DOCX documents and public web page links. The model can automatically parse document and web page content and generate new videos with consistent styles. The generation time is greatly increased to a maximum of 30 seconds, the frame rate is 30fps, lines, BGM and sound effects are natively output, and resolution output of 480P, 720P and 1080P is supported. The bottom layer uses a spatio-temporal joint attention diffusion architecture (Spatio-Temporal DiT) to simultaneously calculate the geometric features of the picture and the inter-frame motion flow in the latent space, improving the stability of long-term videos.

Context

0 tokens

Max output

0 tokens

Input

llm image video

Output

video