Img2Img Alternative Test

[Nargodian]

So this feature (which was the secret sauce to Corridor digital's Anime Rock Paper Scissors) is not present as far as I can tell.

The processes I think noises the image in such a way that the noise sort of maps to the image its self so that when you have similar images like frames of a video, the noise pattern is the same on similar objects. How it dose this I'm not 100% sure, but it would sure be useful to have.

Unless it does already exist which of course will a happy little mistake by me.

[WASasquatch]

I think to achieve this they're just Gaussian noise the tensor based on the mean and standard deviation of the pixel values. So each pixel value gets a relative effect, and similar pixels in similar images get the same effect.

Something along the lines of this perhaps

# Some image tensor
mean = torch.mean(image)
std = torch.std(image)
noise = torch.randn_like(image)
noisy_image = image + std * noise + mean
noisy_image = torch.clamp(noisy_image, 0, 1)

I have dyscalculia, and terrible with math, but what I envision here is the noise of a image sequence (video) would be more or less the same besides the actual moving bits of the frames.

[Nargodian]

So checking over the code in img2imgalt.py of A1111's ui, It seems to be a much more intricate process. But I'm no expert in this kind of stuff.

[WASasquatch]

So checking over the code in img2imgalt.py of A1111's ui, It seems to be a much more intricate process. But I'm no expert in this kind of stuff.

Seems to be using de-noising from the model to compute it's noise, and then that result divided by standard deviation.

Looks like it could be implemented. I am assuming the modes is for 2.x?

[Nargodian]

It works with 1.5, haven't tried it with any 2.x models