NederlandsMiniMax: new image model makes a large, sharp picture in under ten seconds
Chinese AI company MiniMax has released H3-Image, an AI model that creates and edits images. According to MiniMax, it makes a 16-megapixel image in under ten seconds. The model can be downloaded, but its licence does not apply in Europe.
Chinese AI company MiniMax released H3-Image , an AI model that creates images. A user describes in one sentence what they want to see, and the model makes the picture. It can also edit existing images and use up to 16 example images at once, for instance photos of a product that has to appear in a new setting. According to RuntimeWire, MiniMax is mainly aiming at product photos for online shops and advertising.
A large image in ten seconds
The most striking feature is the speed. According to MiniMax, H3-Image makes an image of 4096 by 4096 pixels in under ten seconds. That is about 16 million pixels, as many as a photo from a good camera. The company says it is up to five times as fast as common image models. MiniMax also says the model follows instructions better than other image models that anyone can download. Spectrum Web Co points out that independent testers still have to check these figures.
Available to download, but not in Europe
Anyone with a powerful computer can download the model from Hugging Face, a website where AI developers share their models. The download consists of the model weights: the 33.1 billion numbers in which the model has stored what it has learned.
Together with the part that reads the instruction, the files add up to more than 130 GB. That makes it a model for the powerful computers in data centres, not for a laptop.
For readers in Europe there is an important restriction. The weights come with MiniMax's own licence, and that licence does not apply in the European Union, the United Kingdom, South Korea or the United States. The downloaded model may therefore not be used there. In the rest of the world, too, companies with more than 1 million US dollars in annual revenue, and anyone who offers the model as a service to others, first need written permission from MiniMax. Earlier, Chinese company Baidu also released a large AI model under a licence that does not apply in the EU, while the model can simply be downloaded elsewhere.
Those who do not want to run the model themselves can use it through MiniMax's online services, such as its platform for developers and the Hailuo AI website. There is a standard mode that, according to MiniMax, is 1.5 to 2 times as fast as common services, and a fast mode that is 3 to 4 times as fast. The company says the price is 50 to 70 per cent of comparable services.
What does this mean for the road to the singularity?
AI models that create images are getting faster and cheaper. H3-Image is a large image model that a company does not only offer as a service, but also makes public, so that others can build on it. If MiniMax's figures are right, sharp images for online shops, designers and advertisers will become a lot faster and cheaper to make. The next step is for independent testers to check whether the model really is as fast and as good as the company says.
What is the technological singularity? Read the explanation →
