3 ms·
With a mere moment's thought, one thinks of the obvious way to defeat this: Simply rescale all images in a preprocessing step. Since this defeat is so easy, a
by natch 3y ago
With a mere moment's thought, one thinks of the obvious way to defeat this:
Simply rescale all images in a preprocessing step.
Since this defeat is so easy, and so obvious, either the authors are daft, or they are brilliant and have thought of a way to prevent it.
So, not wanting to waste much time on a paper from daft people, I did a quick skim of the paper. I don't see any mention of this defeat. No prevention of it, no assessment of it, not even any evidence of awareness of it.
Did I miss something?
I'd hate to see technically naive artist types (a mere subset of artist types, I acknowledge) being hoodwinked by such snake oil and paying money for such "tools."
- delecti 3y agoIf you rescale all your training data, then you're training on deliberately blurry images.
- byteknight 3y agoThat's not accurate. Downsizing doesn't always blur.
- natch 3y agoEver so slightly degraded perhaps in some cases but "blurry" is overstating it. And worth the trivial cost if such dumb-as-bricks techniques take hold. One could also submit each image to an image-to-image model with a blank prompt and the right settings to essentially ask for almost a copy, and come out with washed images. Could even do a sharpening (or upscale and sharpening for that matter) at the same time. It has a small (tiny really) GPU cost per image but it's a one-time cost.
- riidom 3y agoYou could use it to check if the image is poisoned. Original image is a "cat", downscaled/blurred/posterized or whatever is a "dog", then something is strange and better skip it.
- natch 3y agoI think “whatever” in your sentence is actually “have an infallibe human who knows all existent artistic styles tag it accurately with subject and style descriptions and then fine tune or train a model with it and then see if the model prompted with those descriptions produces images that match those descriptions.” So I think simply rescaling would be a lot easier.
- numpad0 3y agoIs it legal to willfully circumvent DRM for the purpose of ML training?
- Intralexical 3y agoAs with any DRM technology, the most it can hope to accomplish ultimately is to put more barriers and risks in the way. It won't stop the unwanted behavior completely, but provided it does what it says it shouldn't have zero impact either.