Adobe Is Paying for Videos to Train a New Text-to-Video AI Model: Report

Adobe is reportedly working on building an artificial intelligence (AI)-powered text-to-video generation model. To train its AI model, the company is said to be purchasing videos from photographers and artists. This data will be used in addition to the platform’s existing library of stock images and videos. Interestingly, the software giant is paying an average of $2.62 (roughly Rs. 220) for every minute of submitted videos. Notably, earlier this year, Adobe unveiled its Project Music GenAI Control, an AI music-generation tool which is still under development.

According to a report by Bloomberg, the company has offered its network of photographers and artists as much as $120 (roughly Rs. 10,000) to submit videos, which will then be used to train its AI video model. On the basis of documents seen by the publication, it added that the company has requested videos of “people engaged in everyday actions such as walking or expressing emotions including joy and anger”. It appears Adobe might use this data to train human expressions and natural motion through these procured videos.

Further, the report highlighted that the software giant has requested as many as 100 short clips of people expressing emotions and shots of human anatomy such as feet, hands, or eyes. It also has asked for videos where people are “interacting with objects” such as smartphones and gym equipment.

The documents seen by the publication also warn the collaborators to not submit copyrighted material, nudity or any other offensive content. The payout for this task on average is $2.62 but the report states that some submitters could also be paid as high as $7.25 (roughly Rs. 600) for a minute of submitted video.

The development also highlights the increased costs companies are now incurring for procuring data to train their AI model as the publicly available data sources are being used up. While some tech companies have resorted to ethically procuring the data, others have been accused of stealing copyrighted data from social media platforms. A recent report claimed that OpenAI used more than a million hours of transcribed data from YouTube videos to train GPT-4.