MiniMax H3 Image open source: 4K pictures are produced in about ten seconds, and can feed up to 16 reference pictures.

4096×4096, about 16 million pixels, and it takes about 10 seconds to generate one image. This is the speed given by the newly released image model MiniMax H3 Image from Xiyu Technology, and this time the weights are open.

MiniMax doesn't just put APIs this time around. The model supports three usages: Vincent diagram, image reference and image editing. The weight has been added to Hugging Face, the open platform API has also been opened simultaneously, and its own MiniMax Code and Conch have been connected.

The official said there are three key points for the upgrade: complex command understanding, multi-picture reference, and picture details.

It is most intuitive to refer to multiple pictures. Up to 16 reference pictures can be input at one time. Characters, clothing, products, and scenes can be input separately to allow the model to be assembled into one picture. In the past, when making this kind of drawings, a common practice was to draw a base picture first and then redraw the parts in rounds. Once there were too many reference pictures, the faces or product details that had been determined earlier began to deviate.

Detail is a hard threshold for 4K. After the resolution is increased to 4096, it can be seen at a glance whether the cloth texture, text edges, and skin are blurry. Problems that can be blurred at low resolutions will be exposed at this size.

There is also a Highspeed high-speed version in terms of speed. According to official standards, the generation speed of standard mode is about 1.5 to 2 times that of mainstream tools, and the speed of high-speed mode is 3 to 4 times.

The price is about US$0.069 for a standard version of 4K, and there is a discount in the first week of release, about US$0.034 per picture.

If you don’t want to deploy it yourself or connect to the API, you can now experience H3 Image’s 4K rendering directly on PixPix and PixTV. These two sites are the official partner sites of MiniMax. PixPix focuses on pictures and e-commerce materials, while PixTV is a video creation platform. After finishing the picture, you can move on to the video.

The importance of opening weights will not be clear until the community develops fine-tuned versions, LoRA and workflows. In the first few weeks, we will first look at two things: whether the 16 reference pictures can be stable on real materials, and whether the small text under 4K will collapse.