As the race for AI LLM (large language model) training data heats up further among the major tech giants, the need for fresh content is becoming serious.
This has contributed to Adobe’s decision to buy videos from ordinary users for the sake of training its own text-to-video AI platform.
According to reporting from Bloomberg, Adobe, instead of simply scraping the digital content already posted by others online, will instead take a much more ethical approach to accumulating reams of training videos.
Bloomberg states that the company plans to do this by
“offering its network of photographers and artists $120 to submit videos of people engaged in everyday actions such as walking or expressing emotions including joy and anger, according to documents seen by Bloomberg. The goal is to source assets for artificial intelligence training, the company wrote,”
Given the total amount of money and the duration of clips that Adobe is seeking, this would mean that the company is willing to pay roughly $2.60 per minute of video. Bloomberg however notes that the payments can go as high as $7.25 per minute in some cases.
For generative AI training, Adobe would specifically need certain types of video, mainly those that involve lots of natural human interaction and emotional expression.
Thus, despite the company’s ownership of a huge stock library of over 3 million 4K videos, many of the commercial productions inside that video trove are, ironically, not quite good enough for Adobe’s AI needs.
The very same professional production polish that makes them so useful in commercial contexts also makes them less than ideal for training LLM AI systems to act as naturally human as possible.
Another interesting aspect of this reported effort to buy millions of videos from ordinary users of ordinary human activities is its above-mentioned ethical aspect.
Instead of simply pirating.. sorry.. scraping them from existing online sources, Adobe is directly offering cash incentives to users for voluntary submissions to the company itself.
This closely mirrors what Adobe previously did with its Firefly suite of AI apps, which were trained with content already submitted by users to Adobe and openly used for training the AI behind Firefly.
Some might not agree even with this approach to training image rendering AIs but it’s certainly a lot more open and honest than the notoriously opaque scraping practices of other companies such as Open AI, owner of ChatGPT and DALL·E 3.
To train both its chat AI and its image rendering system, Open AI has reportedly used content scraped from millions of creators who never knew that their creations were being used in this way, or consented to this use.
The same goes for Open AI’s Sora video text-to-video platform Sora, which has also possibly been developed by scraping reams of YouTube videos without their creators’ consent, or Alphabet’s for that matter.
Compared to such practices, Adobe looks positively angelic in its efforts at amassing enough video for the exceptionally difficult job of training an AI that can generate realistic movies and clips from text prompts.
