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?
I compared the effects of StableCascade and standalone VAE under 512x512 sample images, and the results are as follows:
origin image:![image](https://github.com/Stability-AI/StableCascade/assets/12826558/7c853b21-29cb-4fe8-b8e5-2ec410ac17f4)
StableCascade stage B&A:![image](https://github.com/Stability-AI/StableCascade/assets/12826558/ec51cc2a-6a27-4e74-b639-f72140e4d278)
VAE:![image](https://github.com/Stability-AI/StableCascade/assets/12826558/b54192c7-70c4-4dc9-a438-66b4698ec331)
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?