Defending Artists from Fashion Mimicry
What Is Glaze?
Glaze is a device to assist artists to forestall their
creative types from being discovered and mimicked by new AI-art
fashions equivalent to MidJourney, Steady Diffusion and their
variants. It’s a collaboration between the College of
Chicago SAND Lab and members of the skilled artist
group, most notably Karla Ortiz. Glaze has been evaluated
by way of a consumer examine involving over 1,100 skilled artists.
At a excessive stage, here is how Glaze works:
- Suppose we wish to shield artist Karla Ortiz’s
art work in her on-line portfolio from being taken by AI
firms and used to coach fashions that may imitate Karla’s
type. - Our device provides very small adjustments to Karla’s
unique art work earlier than it’s posted on-line. These adjustments
are barely seen to the human eye, which means that the art work
nonetheless seems almost equivalent to the unique, whereas nonetheless
stopping AI fashions from copying Karla’s type. We check with
these added adjustments as a “type cloak” and altered art work as
“cloaked art work.”
For instance, Steady Diffusion at this time can be taught to create
photographs in Karla’s
type after it sees just some items of Karla’s unique
art work (taken from Karla’s on-line portfolio). Nonetheless, if
Karla makes use of our device to cloak her art work, by including tiny
adjustments earlier than posting them on her on-line portfolio, then
Steady Diffusion is not going to be taught Karla’s creative
type. As an alternative, the mannequin will interpret her artwork as a
completely different type (e.g., that of Vincent van
Gogh). Somebody prompting Steady Diffusion to generate
“art work in Karla Ortiz’s type” would as an alternative get
photographs within the type of Van Gogh (or some hybrid). This
protects Karla’s type from being reproduced with out her
consent. You’ll be able to learn our
research paper (at present below
peer assessment).
With Glaze (above), a mannequin trains on cloaked variations of
Karla’s artwork, and learns a distinct type from her unique
visible type. When it’s requested to imitate Karla, it produces
artwork that’s distinctively completely different from Karla’s type.
Some media protection of our undertaking:
- New York Times, by Kashmir Hill
- UChicago CS News, by Rob Mitchum
- TechCrunch, by Natasha Lomas
- Lifewire, by Sascha Brodsky
- Tech Times
- The Register (UK), Katyanna Quach
- Artnet, Jo Lawson-Tancred
- DIY Photography, Alex Baker
- WTTW News, Paul Caine
- La Nacion, Europa Press
- Xataka, Javier Marquez
- Tech Register UK
- NBC News Nigeria
- Heise Online, Martin Holland
- Levtech Japan, Hiroki Yamashita
- Redshark, by Phil Rhodes