What Happened with the VAE? It is Broken!

#2
by qpqpqpqpqpqp - opened

Check https://imgsli.com/NDM1MTI2 it generates artifacts, too sharp and even shifts an image, zoom in to see
Uledxx

it's a great model, and it works for what it was meant to do - HIGH res vae translation.

Yes, there is a 1 pixel shift. easy to fix, I'll publish my node that adjusts it as part of the decode. (yeah, you'll still have the problem of that 1 pixel border on 2 sides, but it'll at least line up with the other VAEs. I suspect it's a rounding problem since the model was trained on 4k, and so lower resolutions end up off by 1.

@scruffynerf But it generates artifacts and extreme sharpness anyway, I don't think UltraFlux is something special, they even have a dedicated repo on github for it. idk what alchemy is it, but the AE is broken, 100% confirmed by me
@Owen777 What do you think? Even if it's intended, I don't think it was broken to train UltraFlux, and I don't get why some decided to start using the autoencoder, who got that it's different and has been promoting it?

@scruffynerf Thank you for the technical insight! The pixel shift may be related to the alignment during 4K training. I appreciate you sharing the workaround/node with the community.

@qpqpqpqpqpqp I noticed you have been opening spam Pull Requests titled "Delete Junk" on multiple repositories (including Vision-CAIR and various community fine-tunes) in the last few hours. Given this pattern of disruptive behavior across the platform, I will not be engaging further with your comments here.

Owen777 changed discussion status to closed

@Owen777 lmao, you are so funny, calling me a spammer due to helping users by making PRs to fix repos. I will tell everyone that you are a troll

Owen777 locked this discussion

Sign up or log in to comment