from Hacker News

Adobe's AI image generator, only trained on licensed or out of copyright content

by fakelvis on 3/21/23, 6:52 PM with 3 comments

• by mrandish on 3/21/23, 8:36 PM

  Adobe + AI is an interesting case of a market-leading incumbent navigating a significant disruption. Adobe saw the potential of AI quite early (~2015) and invested heavily in hiring researchers but hasn't so far shown any ability to harness the disruptive aspects of AI to target new types of customers (ie non-pros).

  I suspect the problem is that Adobe's thinking has been gated by the perspective and concerns of their skilled professional users, thus they've concentrated on AI "features" automating time-consuming or repetitive tasks in current workflows versus entire new "tools" enabling new classes of customers who aren't even interested in the old workflows. It's similar to classic "Innovator's Dilemma" but the twist for Adobe as a creative tooling company isn't "don't disrupt our products", it's "don't disrupt our customer's business."

• by ThrowawayR2 on 3/21/23, 7:13 PM

  This is interesting since it highlights the unfortunate futility of artists' attempts to exclude their own works from being used in AI training. 1) There's more than enough IP unencumbered images out there to build AI image generators anyway, 2) as time moves on copyright expires on more works, the body of images available for training is going to just keep growing, and 3) if a particular artist's style becomes in demand by paying customers, Adobe can easily afford to commission enough low cost artwork that imitates that style and add it to their training corpus¹.

  ¹ Remember that "look and feel" was deemed not copyrightable in https://en.wikipedia.org/wiki/Apple_Computer,_Inc._v._Micros....

• by colesantiago on 3/21/23, 7:07 PM

  Stable Diffusion / Stability AI is going to have a hard time in court trying to defend their model being trained on copyrighted content against Getty Images and other artists.