Has anyone compared the two decoders on faces and fine detail?
Hi everyone, has anyone compared the regular VAE decode and the diffusion decoder using the same generated latents?
I’m interested in old-film restoration, and one question I keep coming back to is how to judge added detail. A face can look sharper in a still frame, but I’d also want to check whether the expression stays consistent and the detail holds up during movement.
If you’ve tried both decoding paths, where did you notice the biggest difference?
Faces, hands, background texture, or flicker between frames?
A short comparison or even a description of what improved—and what got worse—would be helpful.
If possible, include your Diffusers version and whether you used tiled decoding so others can interpret the results.
I’m still figuring out whether this approach fits restoration, so examples where it changes too much would be useful too.