Haiper Ltd., a venture-backed artificial intelligence startup, today debuted a video generator that can create short clips based on user prompts.
Haiper 2.0 is a new iteration of an AI model that the company debuted earlier this year. Compared with the previous release, it generates clips faster and in a more realistic style. Additionally, an upcoming update will boost the resolution of the videos that the model creates to 3840 by 2160 pixels.
London-based Haiper was launched in late 2021 by Chief Executive Officer Yishu Miao and Chief Technology Officer Ziyu Wang, who previously held research roles at Google DeepMind. The company went on to raise a $13.8 million seed round earlier this year. Haiper uses its AI models to power a suite of cloud services that consumers can use to generate short clips.
Haiper 2.0, the new model that the company debuted today, is based on a so-called DiT architecture. This is an approach to building neural networks that was first introduced by Meta Platforms Inc. researchers in a 2022 paper. It combines two earlier neural network designs known as the diffusion and transformer architectures.
To create a diffusion model, developers assemble a collection of images and introduce a type of error known as Gaussian noise into each file. They then instruct the neural network to try to remove the error. By repeating the task many times during training, diffusion models learn how to generate entirely images from scratch.
The DiT design on which Haiper 2.0 is an improved version of the diffusion architecture. To create a DiT model, researchers take a diffusion model and replace several of its key components with a second AI model. The latter algorithm is based on a neural network design known as the transformer architecture, which is most commonly used to build large language models.
The Meta researchers who introduced the concept in 2022 determined that DiT often outperform standard diffusion models. Moreover, algorithms based on the former technology can ingest data in the form of latent spaces. Those are mathematical structures capable of representing files in a condensed, hardware-efficient form that cuts infrastructure costs.
“With our new model, users will be able to generate ultra-realistic videos faster than ever before,” Miao said.
Haiper will use Haiper 2.0 model to enhance its lineup of video generation services. The application suite enables users to generate six-second clips by entering a prompt or uploading a reference image. Additionally, Haiper offers a tool that can extend existing clips by two seconds.
The company’s video generation services face competition from both fellow startups and established industry players. Last week, Adobe Inc. debuted an AI model that will allow customers to edit and extend short clips using natural language prompts. OpenAI, meanwhile, has developed a model called Sora that is capable of generating one-minute videos.