Comment by eigenvalue
3 years ago
This seems like a pretty pointless "arms race" or "cat and mouse game". People who want to train generative image models and who don't care about what artists think about it at all can just do some basic post-processing on the images that is just enough to destroy the very carefully tuned changes this Nightshade algorithm makes. Something like resampling it to slightly lower resolution and then using another super-resolution model on it to upsample it again would probably be able to destroy these subtle tweaks without making a big difference to a human observer.
In the future, my guess is that courts will generally be on the side of artists because of societal pressures, and artists will be able to challenge any image they find and have it sent to yet another ML model that can quickly adjudicate whether the generated image is "too similar" to the artist's style (which would also need to be dissimilar enough from everyone else's style to give a reasonable legal claim in the first place).
Or maybe artists will just give up on trying to monetize the images themselves and focus only on creating physical artifacts, similar to how independent musicians make most of their money nowadays from touring and selling merchandise at shows (plus Patreon). Who knows? It's hard to predict the future when there are such huge fundamental changes that happen so quickly!
>Or maybe artists will just give up on trying to monetize the images themselves and focus only on creating physical artifacts, similar to how independent musicians make most of their money nowadays from touring and selling merchandise at shows (plus Patreon).
As is, art already isn't a sustainable career for most people who can't get a job in industry. The most common monetization is either commissions or hiding extra content behind a pay wall.
To be honest I can see more proverbial "Furry artists" sprouting up in a cynical timeline. I imagine like every other big tech that the 18+ side of this will be clamped down hard by the various powers that be. Which means NSFW stuff will be shielded a bit by the advancement and you either need to find underground training models or go back to an artist. .
>need to find underground training models
It's not particularly that hard. The furry nsfw models are already the most well developed and available models you can get right now. And they are spitting out stuff that is almost indistinguishable from regular art.
> This seems like a pretty pointless "arms race" or "cat and mouse game".
If there is any "point" of this, it's that's going to push the AI models to become better at capturing how humans see things.
> musicians make most of their money nowadays from touring and selling merchandise at shows
Be reminded that this is - and has always been - the mainstream model of the lineages of what have come to be called "traditional" and "Americana" and "Appalachian" music.
The Grateful Dead implemented this model with great finesse, sometimes going out of their way to eschew intellectual property claims over their work, in the belief that such claims only hindered their success (and of course, they eventually formalized this advocacy and named it "The Electronic Frontier Foundation" - it's no coincidence that EFF sprung from deadhead culture).
It is a funny appearance (weird viewpoint) that artists are furious loosing their monopily in stealing and cloning components from other artists, recomposing into a similar but new thing.
And that OpenArt on the analogy of OpenSource is a non-existing thing (I know, I know, different things, source code is not for the generic audience and can be hidden on will, unlike art, just having some generative thoughts artefact here ;) )
the point is you could circumvent one nightshade, but as long as the cat and mouse game continues there can be more