Chinese e-commerce giant Alibaba Group Holding Ltd. has introduced its new artificial intelligence (AI) model, Animate Anyone, which is described as being capable of animating nearly anything from a still image.
Developed by the company’s research team, the Institute for Intelligent Computing, the technology specialized in transforming photos into videos with ‘less glitches and no extra fingers,’ according to Alibaba.
The Hangzhou-based firm also stated that Animate Anyone would soon benefit influencers by generating short-form videos from one image, potentially enabling new use cases such as AI-made outfits and establishing a market for tailored mass-created videos.
Garnering significant interest online, Alibaba plans to offer a demo and code of its new image-to-video method for interested users, with the date yet to be determined.
Animate Anyone Still Needs Work
Alibaba’s Animate Anyone marks an upgrade to previous AI models like Disco and Dreampose.
However, experts have warned that the technology may cause further issues if used irresponsibly.
Image-to-video techniques such as Animate Anyone work by obtaining the necessary details like the subject’s facial features, patterns, and pose in the photo.
A set of images is then created based on those specifics that have been mapped onto slightly different poses, which are later motion-captured or obtained from another video.
Earlier similar models have proven such a capability is feasible, although it was without complications.
Hallucination was reportedly a major issue, as the technology was expected to generate small details, such as how a part of the individual’s clothing or hair might move according to their motion.
The complexity of the process has resulted in oddly generated visuals, leading to videos that were not close to seeming real.
Animate Anyone presents an improved version of the previous AI animation tools, albeit it is a work in progress.
It was noted that Alibaba’s model has a new intermediate step that allows it to learn the connection with the referenced photo in a consistent feature space, considerably helping improve the maintenance of appearance details.
