Alibaba (09988) announces the public testing of its video generation large model Wan 3.0.

On August 6, Alibaba (09988) launched the public beta of its video generation model Wan 3.0. Wan 3.0 has been comprehensively upgraded in dimensions such as generation duration, versatile creation, comprehensive reference, and realism: it can generate a 30-second video in a single instance, fully expressing the creative intent; in addition to the four basic modes of text, images, audio, and video, it now supports document formats like doc, xls, ppt, pdf, and md for the first time; it aims to accurately replicate the real worldeach character appears unique, and every frame is both realistic and believable. Wan 3.0 supports the generation of single 30-second videos, allowing for a more fluid narrative pace, fully expressing complex camera movements such as continuous tracking shots and single-take sequences, transitioning AI video from "generating a shot" to "telling a complete story." It also features an intelligent duration function, which recommends suitable lengths based on prompts; with the video extension feature, it can easily expand the storyline and plot development, allowing creativity to grow naturally. The portrayal of characters in videos generated by Wan 3.0 strives for "thousands of faces," capturing facial features and skin details with greater sensitivity, while the expression of characters' emotions is natural and restrained, with micro-expressions and body movements connected to one another; even in group scenes, it can delicately present the complex emotions of different characters. In comprehensive reference tasks, Wan 3.0 can accurately replicate reference details while maintaining consistency in key dimensions such as characters, props, sounds, spatial relationships, and styles. For characters, fine-grained features such as facial features, hairstyles, body shapes, clothing, and accessories can be consistently maintained to avoid breaking immersion; for props, multi-angle appearances, hardware structures, logos, and material details remain stable; for scenes, it accurately handles the relationship between character positioning and camera angles; for styles, it precisely renders the tone of film shooting, minimizing style confusion during multi-style creations. For digital information, Wan 3.0 enhances the aesthetic level of the visuals, paying attention to harmonious color combinations while maintaining a smooth presentation of information, allowing charts, data, and interface content to possess texture as well. Wan 3.0's video editing capabilities have also significantly improved, supporting modifications to visuals, plot, and dialogue. However, there is still room for improvement in sound quality and textual accuracy.