Recently, Haiper Ltd. launched a video generator capable of automatically creating short videos based on user instructions.
The Haiper 2.0 video generator is an enhanced version of the company's AI model released earlier this year. Compared to its predecessor, Haiper 2.0 offers significantly improved video generation speed and realism. Additionally, upcoming updates plan to increase the model's video resolution to 3840x2160 pixels.
Headquartered in London, Haiper was co-founded by CEO Yishu Miao and CTO Ziyu Wang at the end of 2021. Both previously held research positions at Google DeepMind. Earlier this year, the company successfully raised $13.8 million in seed funding. Haiper leverages AI models to support a range of cloud services, enabling users to create short videos using these services.
The recently released Haiper 2.0 model is based on the DiT architecture, first introduced by researchers at Meta Platforms Inc. in a 2022 paper. The DiT architecture combines diffusion models and transformer neural network designs.
The process of constructing diffusion models involves collecting a large set of images and introducing a type of noise called Gaussian noise to each image. The neural network is then trained to eliminate this noise. By repeatedly performing this task during training, the diffusion model learns how to generate complete images from scratch.
The DiT design underpinning Haiper 2.0 is an improved version of the traditional diffusion architecture. When building a DiT model, researchers start with a diffusion model and replace several key components with a second AI model based on the transformer architecture, which is commonly used to develop large language models.
Meta's research team, which introduced the DiT concept in 2022, found that DiT models generally outperform standard diffusion models. Additionally, algorithms based on DiT can ingest data in a latent space form. Latent space is a mathematical structure that represents files in a compact and hardware-efficient manner, helping to reduce infrastructure costs.
Haiper plans to leverage the Haiper 2.0 model to enhance its suite of video generation services. This application suite allows users to create six-second short videos by inputting instructions or uploading reference images. Additionally, Haiper offers a tool that can extend the duration of existing videos by two seconds.
