The AI-powered video generation landscape is gaining momentum with the recent release of Haiper by two DeepMind alumni, Yishu Miao and Ziyu Wang. This comes on the heels of OpenAI’s Sora model, contributing to the growing interest in innovative video synthesis technologies.
Yishu Miao, formerly part of TikTok’s Global Trust & Safety team, and Ziyu Wang, a research scientist with experience at both DeepMind and Google, began working on Haiper in 2021, formally incorporating the company in 2022.
Their expertise in machine learning initially led them to explore 3D reconstruction using neural networks. However, after delving into video data training, they shifted their focus to the captivating challenge of video generation approximately six months ago.
In a notable achievement, Haiper secured $13.8 million in a seed round led by Octopus Ventures, with participation from 5Y Capital. Prior to this, angel investors like Geoffrey Hinton and Nando de Freitas supported the company with a $5.4 million pre-seed round in April 2022.
Haiper’s Video Generation Service
Haiper’s user-friendly platform allows users to generate videos for free by entering text prompts. However, there are limitations on video length, with a maximum of two seconds for HD videos and slightly lower quality for four-second videos.
The platform offers additional features such as image animation, video restyling, and ongoing efforts to introduce capabilities like video extension.
Miao emphasizes the company’s commitment to keeping these features free to foster a community. Although it has engaged in collaborations with companies like JD.com for potential commercial applications, the primary focus remains on community building.
Future Plans and Core Video Model
While Haiper currently concentrates on its consumer-facing website, the long-term goal is to develop a core video generation model that could be shared with others. The specifics of this model have not been disclosed, but Haiper has privately reached out to developers to test its closed API.
Miao stresses the significance of developer feedback, indicating rapid iterations on the model. While considering the possibility of open-sourcing its models in the future, Haiper aims to address fundamental challenges, particularly the “uncanny valley” problem in video generation.
Competition Landscape
In a field teeming with competition, Haiper faces notable rivals such as OpenAI’s Sora, Google and Nvidia-backed Runway, Google and Meta with their own video generation models, and Stability AI’s Stable Diffusion Video model. The challenge for Haiper lies in differentiating itself by building a robust video generation model that transcends the uncanny valley.
Investors, including Octopus Ventures, recognize the potential of AI-powered video generation but acknowledge the need for further advancements. Rebecca Hunt, a partner at Octopus Ventures, anticipates a pivotal moment when AI video models reach a level that reflects the real world and its physics, unlocking endless applications.
As Haiper actively hires across engineering and marketing roles, the industry awaits developments in this dynamic space. Investors and experts anticipate continued improvements in AI-powered video generation, eagerly awaiting a breakthrough akin to the “ChatGPT moment” for video technology.