Topaz Labs, which produces photo and video improvement software, introduced the new artificial intelligence model that improves old and low -resolution videos. In the sample videos published by the company, it is seen that the model significantly increases the details and greatly reduces the defects in the video.

FIRST VIDEO BUILDING BASED VIDEO IMPROVING MODEL

The company says that the project Starlight is the first diffusion -based model developed for video restoration and its latest Nvidia hardware with a new model architecture with more than 6 billion parameters. In terms of comparison, OpenAI’s GPT-4O large language model requires 8 billion parameters.

Topaz Labs promises that the model can correctly reconstruct the details in the video and offer unique timely consistency with unique detail recovery. The company states that this is exactly the main focus of the model: by improving multiple frames, to achieve high quality restoration results without motion deterioration or inconsistencies between squares and objects.

How do diffusion models improve the video?

Diffusion models gradually add noise to it by looking at a high -quality image and learn how this process disrupts the image. It also predicts how it looks before the noise is added by doing the opposite and starting from a noisy image.

Project Starlight is also capable of automatically relieving noise, turbidity reduction, increasing resolution and edge softening. In summary, it can automatically perform what professional programs do.

The video restoration includes various processes such as scaling, color correction and rating, square enterpolation, damage repair and sound restoration. In order to help each of these, artificial intelligence -supported vehicles are bouquet, but at the moment, people need to control everything to achieve the best results. Project Starlight automatically performs these operations.

When we look at the sample videos, we can say that Project Starlight mostly does a very good job. The clip with astronaut and red parrot looks especially good. However, the detail accuracy on the boxing match scene does not meet the expectations. Some frames are similar to blurred pictures seen in old artificial intelligence video production models.

Topaz Labs says that videos up to 10 seconds can be restored for free. Clips up to 5 minutes can be processed in a maximum resolution of 1080p and require the use of loans. The corporate version will support longer videos and higher resolution outputs.