Topaz Labs, specializing in the development of software for editing photos and videos, introduced the model of AI Project Starlight for improving the quality of old frames from home video collection or archival content, the quality of which could have worsen over time during the storage of traditional media.
Image source: Topaz Labs
According to the developer, this is the first diffusion model in history, created for these purposes, and it does not require manual data entry to correct the video. It is reported that the Project Starlight was created from scratch using the new architecture of the model with more than 6 billion parameters, and its work is supported by advanced NVIDIA accelerators. For comparison, published in May 2024, the GPT-4O Language Model from Openai with the possibility of processing text, audio, images and video as input data initially had 8 billion parameters.
Topaz Labs claims that the model “accurately restores the details” and provides “unsurpassed restoration of parts in combination with unsurpassed temporary consistency”. According to the company, this is precisely the essence of its new model: improving several frames to achieve high -quality restoration results without artifacts of movement or inconsistencies between personnel and objects.
Project Starlight also automatically removes noise, eliminates blur, scale and smoothes frames on request. To work with this AI model does not at all require special knowledge in the field of video processing.
The return of the old video to life includes several processes, including scaling, color correction and sorting, personnel interpolation, elimination of damage and recovery of sound. For each of these recovery options, tools based on AI have already been created, but in order to achieve the best results, people should currently be controlled by people.
Topaz Labs said that users can, using its II model, restore a video up to 10 s for free, while clips up to 5 minutes will have the maximum resolution of 1080p and this will require loans. The version for corporate users supports the restoration of longer videos and with a higher resolution. It is not yet known whether the Project Starlight will work locally or will be integrated into other applications of the company.