huggingface / huggingface/diffusers

Attention masks are missing in SD3 to mask out text padding tokens

Open
#8,673 9 comments 0 reactions 0 assignees View on GitHub
contributions-welcome wip
Dominant language
Python
Stars
34.5k
Forks
7.3k
Avg merge
3d 3h
Merged PRs (30d)
91

Description

### Describe the bug

In the attention implementation of SD3, attention masks currently are not used. This will result in inconsistent outputs for the different values `max_seq_length` where padding exists in text tokens as the attention scores of padding tokens are non-zero. This issue has been discussed in https://github.com/huggingface/diffusers/discussions/8628, and is created to track the progress of fixing this problem.

Thanks @sayakpaul for the discussion.

### Reproduction

n/a

### Logs

_No response_

### System Info

n/a

### Who can help?

_No response_

Contributor guide

Open the contributing guide

Research direction

Start with the SD3 attention implementation and read the linked discussion for the existing context on padding tokens. Trace how attention scores are computed when text padding is present, then verify that outputs no longer vary with max_seq_length solely because of padding.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
machine-learning
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.