Regarding the artifacts at the VAE level
Opened this issue · 4 comments
Not sure what your training process is however were you training with the actual 1.0 vae?
or did you use the 0.9 vae? - it's known that an "invisible" (quite visible) watermark is baked into the 1.0 vae. (this is in your relation to the artifacts found with the SDXL VAE)
I trained on images encoded with the v0.9 one in FP32 mode.
I think the encoder for v1.0 and v0.9 are the same, though. This is more to do with the fact that the SDXL vae is a lot more sensitive to any kind of changes in the latents, and my amateurish neural network isn't good enough to actually output something close enough to what is expected.
I've been trying to learn about better NN architectures but yeah, I'm a sysadmin doing this as a hobby. The neural network for this was literal trial-and-error.
Hey, don't put yourself down, this is pretty epic and deffo produces some super nice results. Excited to see any future releases too, deffo an invaluable tool.
Just curious - would you be interested in access to multiple A100's for development?
Just asking because you said it's undertrained and we'd like to help.
multiple A100's for development
I'd be happy just getting access to one :P
Jokes aside, I'd like to ask some follow-up questions privately first if you don't mind. Just boring stuff like location/hardware specs/availability/privacy policy/terms of service/etc. Assuming I didn't mess up my MX records too badly, you can shoot me an email on city@eruruu.net - unless you prefer some other form of communication.