A lot of artists are worried about their uploaded work being automatically collected and used to train generative models without permission. One possible solution would be for an art-sharing website to apply a subtle, specialized noise pattern or adversarial filter to every uploaded image. Humans would still see the artwork normally, but image-recognition and scraping systems would have a harder time reading or training on it. It probably would not be foolproof, especially as models improve, but could it still be a useful layer of protection? Has anyone already built a platform that applies this kind of protection automatically?
5 Answers
There are also browser-based approaches that prevent ordinary text from existing as selectable page content, which can interfere with simple scraping and copying. Those methods can make casual automated collection harder, but they do not stop screenshots or someone running their own optical-recognition system. At best, this is one layer in a broader set of defenses.
Some art platforms have offered similar filters, but they are often opt-in because artists may not want visible artifacts and the processing can be costly at scale. There is also a consent problem that image filters cannot completely solve: if a site allows public viewing, someone can still download, screenshot, or photograph the work.
This is already a known area of research. Tools such as Glaze and Nightshade try to interfere with how models interpret or learn from artwork. The main challenge is that the protection may need to be updated whenever models develop ways to remove or ignore the noise, and processing every upload could also affect image quality or site performance.
Even if it is not permanent protection, making automated collection more expensive and difficult could still be worthwhile as an optional feature.
A filter can be a useful deterrent, but it should not be advertised as a complete barrier. Determined people can often obtain the original pixels, and future systems may be trained specifically to reverse or bypass current protections. Rate limits, bot detection, restricted bulk downloads, provider blocking, and clear licensing rules would probably work better when combined with image-level filtering.
Another perspective is that no public website can guarantee that its images will never be scanned. The realistic goal is to reduce unauthorized collection, document consent and usage rules, and make abuse costly enough that casual scrapers move on. Artists who want maximum control may need to host previews while keeping high-resolution originals private.
That is the frustrating tradeoff. Keeping everything behind a login gives stronger protection, but it also makes it much harder for new people to discover the artist.