Could someone ELI5? What is the impact of this issue?
replies(1):
At the core of a latent diffusion model is a de-noising process. It takes a noisy image and predicts what is noise vs what is the real image without noise. You use this to remove a bit of noise from the image and repeat to iteratively denoise an image.
You can use this to generate entirely new images by just starting with complete random noise and denoising til you get a 'proper' image. Obviously this would not give you any control over what you generated. So you incorporate 'guidance' which controls how the denoise works. For stable diffusion this guidance comes from a different neural network called CLIP (https://openai.com/research/clip) which can take some text and produce a numerical representation of it that can be correlated to an image of what the text describes (I won't go into more detail here as it's not really relevant to the VAE).
The problem you have with the denoising process is the larger the image you want to denoise the bigger the model you need, and even at a modest 512x512 (the native resolution of stable diffusion) training the model is far too expensive.
This is where the latent bit comes in. Rather than train your model on a 512x512x3 representation (3 channels R,G,B per pixel) use a compressed representation that is 64x64x4, significantly smaller than the uncompressed image and thus requiring a significantly smaller denoising model. This 64x64x4 representation is known as the 'latent' and it is said to be in a 'latent space'.
How do we produce the latent representation? A VAE, a variational autoencoder, yet another neural network. You train an encoder and decoder together to encode an image to the 64x64x4 space and decode it back to 512x512x3 with as little loss as possible.
The issue pointed out here is the VAE for stable diffusion has a flaw, it seems to put global information in a particular point of the image (to a crude approximation it might store information like 'green is the dominant colour of this image' in that point). So if you touch that point in the latent you effect the entire image.
This is bad because the denoising network is constructed in such a way that it expects that points close in the latent only effect other points close in the latent. When that's not the case it ends up 'wasting' a bunch of the network on extracting that global data from that point and fanning it out to the rest of the image (as the entire image needs to know it to denoise correctly).
So without this flaw it may be the stable diffusion denoising model could be more effective as it doesn't need to work hard to work around the flaw.
Edit: Pressed enter too early, post is now complete.