init_image support about aphantasia HOT 4 CLOSED

all three main methods can start from an image - fft, dwt, rgb.
this feature is very hacky [and kinda broken] though, in line with my understanding of the functions from the lucent lib. the colors usually become seriously distorted on the load; so i would definitely appreciate any hints regarding fixing this.

from aphantasia.

Thanks - I've found this in the code via the "resume" argument. Agreed the current implementation is a bit rough around the edges, but a useful baseline we might be able to improve over time.

from aphantasia.

I was looking for a way to initialize fft from an image, for a setup that uses FFT for image generation and VGG19 style transfer and CLIP for guiding the iterations. My FFT implementation was borrowed from here, and I had to rework it (largely by trial and error) to get it to start from a good reproduction of the init image.

You'll find my code here https://github.com/htoyryla/fftstylediver

It will probably not work correctly on the latest pytorch versions (have not tested it and I have changed the FFT dimension thing). I have also done some scaling things differently, and dropped to_valid_rgb from the chain.

from aphantasia.

just to follow up on this:
rgb [video] method now properly starts from image. "properly" means 100% empirical hack, but looking quite ok.
fft and dwt are better than before, but colors are washed out on start yet.

from aphantasia.