Haiper, Newcomer in Video AI, Launches Trial to Challenge OpenAI's Sora

Despite the caution we exercise when encountering text and images on the internet with the advent of generative AI, video generation technology is also catching up. Following OpenAI's Sora video creation AI, a startup founded by former Google DeepMind researchers has announced its own video AI called Haiper. Unlike Sora, Haiper is now available for public trial, but it can currently only generate a few seconds of video.

Yishu Miao and Ziyu Wang started developing Haiper in 2021 and established the company in 2022. Initially, they were interested in using AI for 3D reconstruction. However, the process of training AI on video data fascinated them, so about six months ago, they decided to shift their focus to video generation.

The progress of this company has been made public. Although the video clips generated by Haiper may not be as impressive as the samples released by OpenAI, most people have not had the opportunity to experience Sora yet. With Haiper, users can create a two-second HD video based on text prompts or generate a longer clip with slightly lower resolution. Additionally, Haiper can also convert static images into animated versions or "redraw" videos in different styles. The company also plans to add the ability to AI-enhance videos.

Currently, it takes about a minute to generate each sample video, depending on the computing power available to Haiper. All generative AI systems require expensive AI accelerators, such as the Nvidia H100, which can cost over $30,000 per card. As the company grows, it is expected that the video generation speed will improve.

Mao stated that he plans to keep these basic features free, but it is too early to discuss creating a paid service layer. The company has already started reaching out to developers to test its closed API, as this feedback is crucial for improving the model before wide release. The company's leadership is even considering open-sourcing the model for other teams to explore different applications of video generators.

Sora is not the first AI capable of creating videos, but its impressive results and the reputation of OpenAI have garnered public attention. We can anticipate a wave of AI video hype similar to what text and image generation models experienced in the past year. While it is a more complex problem, many companies have raised funds to seek solutions.