Some issues with the effectiveness of image reconstructions

[HelloClyde]

As you can see, the reconstructions are surprisingly close, even for small details. Such reconstructions are not possible with a standard VAE etc.

I compared the effects of StableCascade and standalone VAE under 512x512 sample images, and the results are as follows:

origin image：

StableCascade stage B&A：

VAE：

It seems that a standard VAE can achieve this kind of reconstruction effect. Does the readme imply that a standard VAE cannot achieve this effect under a 24x24 latent space?