Adobe will fork over cash for clips to train text-to-video AI
Not touching copyrighted material with a barge pole
Adobe is building its own AI model capable of transforming text into video and, unlike other companies, will actually pay creators of the material used to train it.
Using AI to generate video from text seems to be the next big thing in AI content creation, with OpenAI taking the lead with its Sora AI model. But whereas OpenAI has often sourced AI training materials from any public sources possible (and has been accused in lawsuits of allegedly including copyrighted works), Adobe is reportedly going down a different path, sourcing footage for training its upcoming AI model by purchasing it from certain filmmakers.
The videos Adobe is requesting are supposed to show pretty mundane things, such as using a phone, people expressing emotions, movements of the human body, and presumably more as Bloomberg mentions Adobe is purchasing 100 different types of footage. The report calculates that compensation for submitted footage ranges from $2.62 to $7.25 per minute, with each video being worth $120.
It's not clear who would be eligible to submit this footage to Adobe, or if there's a limit to how much footage can be submitted. If one person can submit at least 100 clips for $120 each, that would come out to a decent amount of compensation for filming everyday occurrences.
- Canva acquires Affinity, further wounding a regulator-bruised Adobe
- Adobe has 'no plans' to invest in XD despite failed Figma buy
- Former Adobe software engineering leader convicted of insider trading
- Adobe ditches $20B Figma takeover under pressure from monopoly cops
Additionally, there are bound to be at least a few rules submissions have to abide by. Bloomberg says no footage should be offensive, show nudity, or be a copyrighted work, which is standard for Adobe. When the company ran a similar program to train its Firefly AI image generator, it didn't accept photos of brands, public figures, and even specific words per Bloomberg – similar rules could apply to Adobe's program for training its video generator.
Copyright, however, is likely a chief concern for Adobe. OpenAI is embroiled in countless lawsuits over ChatGPT generating summaries of copyrighted novels, thanks to allegedly being trained on said novels. In an interview with the Wall Street Journal last month, OpenAI CTO Mira Murati said she didn't even know whether Sora was trained on YouTube videos, something that YouTube CEO Neal Mohan says would violate the platform's terms of service.
Legally, it's not settled whether it's permissible to train an AI model using copyrighted works. Plus, while Adobe's scheme might be expensive, it could allow the company to source higher quality footage and tweak the upcoming video generation tool to Adobe's liking. ®