lllyasviel / lllyasviel/sd-forge-layerdiffuse

subtraction in attention sharing mechanism

Open
#108 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
4.1k
Forks
352
PR merge metrics
No merged PRs in 30d

Description

In the implementation of attention sharing, I noticed there's a stacked temporal attention adapter.

My question is, why did you subtract the `modified_hidden_states` with input `h`? Could you share some insights behind the rationale of this design? Thanks!

https://github.com/layerdiffusion/sd-forge-layerdiffuse/blob/e4d5060e05c7b4337a3258bb03c4e3ad2f8b15bb/lib_layerdiffusion/attention_sharing.py#L131-L137

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start with lib_layerdiffusion/attention_sharing.py at lines 131-137 and trace the stacked temporal attention adapter's inputs and outputs. Read the surrounding implementation to determine the rationale for subtracting modified_hidden_states from h. Done means documenting a clear explanation of that design choice.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
machine-learning
Issue type
Documentation
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Needs clarification
Newbie friendliness
20/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.