Alibaba is rolling out an update to its video generation model HappyHorse, and this time it's all about keeping things consistent. Version 1.1, now available on Alibaba Cloud's Bailian platform, Tongyi Qianwen, and other services, directly addresses the video generation woes that plagued the 1.0 release — dynamic instability, subject drift, and inconsistent visual quality. The new model promises better instruction following, improved stability, and an overall bump in generation quality.
https://twitter.com/HappyHorseATH/status/2068941160584974753
The headline feature? Support for up to nine reference images. You can feed the model multiple photos of a person, product, or scene to lock in consistent representation across generated clips. This makes HappyHorse 1.1 especially suited for the kind of content that demands stable character output: short-form dramas (a booming format in China known as duanju), e-commerce product videos, and advertising.
On the technical side, the model is accessible via API on Bailian and select third-party platforms, churning out 1080p video with synchronized audio and lip-sync capabilities. Alibaba has also kicked off the HorsePower AI Cinema Awards, a community contest to spur creative use of the model, and released a co-created short film titled Are You Happy to showcase what's possible.