Skip to content

Conversation

@Acly
Copy link
Owner

@Acly Acly commented Oct 16, 2025

No description provided.

Acly added 24 commits October 1, 2025 23:34
* CWHN is a bit faster on CPU (900ms -> 750ms for ViT-S)
* if any of the spatial dimensions are 1, align-corners would run int division-by-zero or oob access
* updated birefnet gpu images after the fix
* added reference images for depth-anything tests
* too much work to migrate atm
* redoing the measurements yields higher values than before for pytorch, not 100% sure if previous times were incorrect
* comparison was against official BiRefNet repo, a3bb3efe2f824ec66644ca5941583c4c90c6e027
* tried with SDPA on/off
* env VISP_FLASH_ATTENTION=0 always disables it
* env VISP_FLASH_ATTENTION=1 always enabled it
* all other values use default
@Acly Acly merged commit d381eaf into main Oct 16, 2025
3 checks passed
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.

2 participants