-
Notifications
You must be signed in to change notification settings - Fork 5.4k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Flux latents fix #9929
base: main
Are you sure you want to change the base?
Flux latents fix #9929
Conversation
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
this seems really important but considering how slowly things move around here, not so sure it will be merged in a timely manner. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
oh thanks for the fix!
I made a comment
width = int(width) // self.vae_scale_factor | ||
# VAE applies 8x compression on images but we must also account for packing which requires | ||
# latent height and width to be divisible by 2. | ||
height = int(height) // self.vae_scale_factor - ((int(height) // self.vae_scale_factor) % 2) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
is this same as this? (basically revert back to the original code but divisible by self.vae_scale_factor*2
a little bit easier to understand, no?
height = 2 * (int(height) // (self.vae_scale_factor*2))
What does this PR do?
This PR: #9711 changed the latent preparation step so that the
vae_scale_factor
and default sample size are more clear.However, we overlooked the fact that this operation (with scale_factor == 16)
Is not equivalent to this operation with scale_factor == 8
For resolutions that are divisible by 8 but not by 16 the current pipeline on main will error out, because the implementation does not account for the fact that the latent width and height must be divisible by 2 for the packing step to work.
The old implementation scaled the height and width by
//16
and then multiplied by 2. This results in slightly resized images if the height or width aren't multiples of 16.This PR:
Fixes # (issue)
Before submitting
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.