© The Indian Express Pvt Ltd
Tags:
Journalism of Courage
The race to bring out the best text-to-video AI model seems to be intensifying. On Tuesday (March 5), Haiper — a company founded by former members of Google DeepMind, TikTok, and top labs from research academia — introduced an eponymous new text-to-video model. The announcement came just weeks after the launch of OpenAI’s AI model Sora, which is capable of generating one-minute high-definition videos from text months.
According to Haiper, its model is a powerful perceptual foundation model-driven AI that has been designed for a “new path towards AGI” (artificial general intelligence, an AI software with human-like intelligence and the ability to self-learn).
Haiper is an all-in-one visual foundation model that allows everyone, with or without technical training, to generate high-quality video content with ease. The founders claim that Haiper brings forward cutting-edge machine learning with the belief that creativity should be “fun, surprising, and shareable”. The company has built Haiper as a powerful, industry-agnostic creativity tool.
Released by DeepMind alumni, Yishu Miao and Ziyu Wang, on Tuesday, the model comes with the company’s own proprietary AI model underneath. According to reports, Miao had previously worked at TikTok’s Global Trust & Safety team, while Wang worked as a research scientist for Google and DeepMind. Both Miao and Wang also have PhDs in Machine Learning from Oxford University. The duo reportedly began working on their venture sometime in 2021 in London and it was formally incorporated in 2022.
Haiper offers tools such as text-to-video, animated static images, video repainting tools, etc. Users can go on to the website, log in with their email addresses, and start generating videos for free by typing in text prompts. At present, users can only generate HD video spanning 2 seconds, and a slightly lower-quality video could go up to four seconds.
To try the model, The Indian Express used the prompt: “Dog using an iPhone while surfing in the desert” in HD settings. The tool created a fairly accurate two-second video of the subject.
At a glance, the output seemed perfect, as a bulldog is seen seated on a skateboard, with an iPhone hoisted vertically in front of the canine. The dog is also wearing a pair of sunglasses. Overall the image seems like a hyper-realistic version of the prompt. However, a closer inspection revealed some not-so-endearing details such as the tail of the dog seems to be melting onto the skateboard, and the sand next to the subject seems like creased fabric. The dog’s paws and collar seemed near perfect.
While the short length is a limitation, the company has said that it is working towards extending the video outputs. Presently, the company has kept the tool free to use, with an aim to build a community. Miao reportedly said that it was too early for the startup to think about working on a subscription model for its product on video generation.
As of now, Haiper is focussed on building a consumer-centric website. Essentially, the founders want to build a core video-generation model that could be offered to various stakeholders. So far, the company has not divulged any details about the model it has been using. Reportedly, the company has reached out to a bunch of developers to try Haiper’s closed API (Application Programming Interface) and is looking forward to feedback. The founders have plans to develop and release a series of large trained models in the next few months.
While OpenAI’s Sora is still not available for the public, Haiper is offering users to try its tool for free on its website. However, the newly found company has a long way to go, especially when it comes to competition considering the dominance of OpenAI and Google.
