Skip to content

Conversation

@sayakpaul
Copy link
Member

@sayakpaul sayakpaul commented Jul 21, 2024

What does this PR do?

Fixes #8887 and #8708.

Additionally, it adds an option to control the pre-conditioning behavior on the model outputs. Multiple folks have reported that for rectified-flows we are better off without it.

@yiyixuxu for awareness.

@bghira FYI.

Pinging @haofanwang for awareness as well.

Co-authored-by: bghira <[email protected]>
@sayakpaul sayakpaul requested a review from kashif July 21, 2024 02:53
@bghira
Copy link
Contributor

bghira commented Jul 21, 2024

if you wanted to rewrite the formula to match that in the paper i think it's been requested and done on downstream implementations. it just reduces confusion when reading. i'm referring to the order of this statement:

                    noisy_latents = (1.0 - sigmas) * latents + sigmas * noise
                    # is equal to:
                    # zt = (1 - texp) * x + texp * z1

@sayakpaul
Copy link
Member Author

Done 9050093

@HuggingFaceDocBuilderDev

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.

@sayakpaul sayakpaul merged commit 1a8b3c2 into main Jul 21, 2024
@sayakpaul sayakpaul deleted the fix-sd3-training branch July 22, 2024 02:56
@wangqixun
Copy link
Contributor

👍👍👍👍

sayakpaul added a commit that referenced this pull request Dec 23, 2024
* SD3 training fixes

Co-authored-by: bghira <[email protected]>

* rewrite noise addition part to respect the eqn.

* styler

* Update examples/dreambooth/README_sd3.md

Co-authored-by: Kashif Rasul <[email protected]>

---------

Co-authored-by: bghira <[email protected]>
Co-authored-by: Kashif Rasul <[email protected]>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

SD3 wrong VAE encode in training script

6 participants