(Credit: Shutterstock/Mats Wiklund)
Adobe is building its own video-generating AI model like OpenAI's Sora. The company behind Photoshop and Premiere Pro software is now offering artists roughly $3 per minute of video footage to train its new AI model, Bloomberg reports, citing documents detailing the plans.
The new model will be text-to-video, like Sora, which ingests a text description and spits out videos up to 60 seconds long. It's unclear whether Adobe's model will be able to create videos a minute in length, or whether its abilities will be more or less.
Adobe is asking its network of artists for video clips, according to the report, promising potential payouts of just $120, which Bloomberg estimates equates to anywhere from $2.62 to $7.25 per minute of video footage. In the past, Adobe has reportedly paid artists roughly six to 16 cents per still image to feed into its AI models.
Adobe's rates aren't great numbers considering the video footage would be used to train an AI model which, like nearly all other AI models in development, don't cite or credit artists for their work. It's also unlikely that artists would see additional licensing revenue or royalties after the initial payments, as such practices aren't yet standard in the world of large language model (LLM) training.
There's also been ongoing concern from artists that AI models may impact future job prospects. Others who have seen their names added to Midjourney image prompts, for example, are upset because they believe their names and artistic styles are being used without their consent.
Adobe wants "more than 100 short clips" of humans showing various emotions, completing different physical activities or doing things with objects, or clips that show the human anatomy. Adobe doesn't want artists to send in any work that's copyrighted, however, and it doesn't want "offensive content" or nudity either, per the report.
100 clips is a low number considering AI models—like some of the smaller models made by Google—are trained on upwards of 500 billion tokens of data. While data "tokens" are a smaller, different metric than a video clip, it's likely Adobe will need substantially more video data to effectively develop a video generator capable of making realistic videos. That said, it's likely Adobe already has a large trove of data to pull from considering its range of existing AI tools, from its text-based video editor to its Firefly image generation tool.
It's unclear how or whether Adobe will investigate each video clip submitted, however, to verify that it doesn't belong to another artist.
Adobe confirmed to Bloomberg that it's developing an AI-powered video generator, but didn't provide further comment. PCMag has also reached out to Adobe for comment. The software firm is expected to share more details about its AI video generator sometime later this year.