4 ms·
Having spent some passionate time with StableDiffusion I can outline a few signs in the details of images this article has missed. - tiny details never make se
by smusamashah 3y ago
Having spent some passionate time with StableDiffusion I can outline a few signs in the details of images this article has missed.
- tiny details never make sense. In real art or photos, no line or edge is useless or senseless. But since generated images are only hallucinations, the tiny lines often blend into something else, convert to something else, or simply fade out.
- scaling is wrong. Most of the times scaling won't make sense. Take a tiny part of the image, just looking at that part, can you guess how big or small other things in that image should be? A box lying on the floor looks way too small or big for example.
- perspective is off. That sponge + cheese image for example the angle of eyes or head or the appearant thickness on that angle isn't right.
Models are getting really good though. Previously you could look at texture of clothes and tell, that is almost gone these and you have to find some other pattern which is usually some other non-symmetry.
Tiny details, textures is the most visible and almost always recognisable sign of a generated (unaltered) image.
- userbinator 3y agoTiny details, textures is the most visible and almost always recognisable sign of a generated (unaltered) image. In some ways, I think AI art resembles what you'd see in a dream. The "big picture" is there and logical, but otherwise it's full of little inconsistencies that don't make sense.
- golergka 3y ago> In real art or photos, no line or edge is useless or senseless. In real art, artists screw up little details all the time. I've recently seen an X thread (unfortunately no longer have the link) where the author quite convisiningly highlighted all the kinds of typical AI artefacts found in the works of Rembrandt.
- smusamashah 3y agoI am not an artist but can draw a little and if by screwing up little details you mean filling the image with random strokes which make the bigure picture look correct than those are not what I am talking about. In generated image when you follow a tiny detail it will just disappear into something else. In painting you can usually see the strokes up close but in a gen image, those strokes will be a weird blend if they are visible. Tiny details like paint smidges etc won't be there. Shadows of tiny details won't make sense. I really like to see that X thread.