3 ms·
I'm mostly just shocked that the model can produce what appear to be entirely consistent models of the same character from different views. I can be convinced,
by crazygringo 4y ago
I'm mostly just shocked that the model can produce what appear to be entirely consistent models of the same character from different views. I can be convinced, but I was also equally skeptical.
Because that implies a kind of cohesive conceptual understanding of a single object, that I didn't think the current architecture was able to model that. So many image outputs of humans are perfectly fine in local areas, but then they have three arms or a hand where a foot should be or something.
I mean, the chosen character isn't perfectly consistent -- e.g. the side view shows a kind of metal badge on the vest that doesn't exist in the front view, while the back view shows a kind of mid-thigh holster (?) that doesn't appear on the front view, plus the vest on the back view is a few inches shorter. Similarly the collar is popped on the front view, but not the other two.
But nevertheless, it's basically about 99% right. And I simply don't understand how that's possible, unless there author did about ten thousand tries and is showing the only ones that happened to be nearly entirely consistent, somehow just mostly by chance. (Maybe because they're particularly close to the real-life characters used for training?)
- minxomat 4y ago> kind of cohesive conceptual understanding of a single object It doesn't, it just requires character sheets in the training set, the rest is basically a style transfer.
- YeGoblynQueenne 4y ago>> Because that implies a kind of cohesive conceptual understanding of a single object, that I didn't think the current architecture was able to model that. It works because it's all in one image, you still can't get the same consistency across different images. Someone did the same thing with sprite sheets the other day: https://www.reddit.com/r/StableDiffusion/comments/yj1kbi/ive_trained_a_new_model_to_output_pixel_art/ https://www.reddit.com/r/StableDiffusion/comments/yj1kbi/ive... Which again works because you're asking for a single image that is similar to other single images (i.e. other sprite sheets). I thought that was very clever. So I guess the next big thing will be to train on datasets of entire graphic novels where all the pages of each novel are stitched together into "one image". Then we will finally achieve the graphic singularity and comic book artists will have been replaced, too :P