Skip to content

Releases: dbolya/tomesd

v0.1.3

14 May 22:18
Compare
Choose a tag to compare

v0.1.3

  • [2023.04.24] Random perturbations now use a separate rng so it doesn't affect the rest of the diffusion process. Thanks @alihassanijr! This makes it so that tomesd has no effect on the underlying torch seed. In platforms that set the seed automatically, this won't have much of an effect, but for those that aren't as careful at setting seeds this will improve results.
  • [2023.04.25] Fixed an issue with the separate rng on mps devices. (Fixes #27)
  • [2023.05.14] Added fallback to CPU for non-supported devices for the separate rng generator.
  • [2023.05.14] Defined use_ada_layer_norm_zero just in case for older diffuser versions. (Fixes #20)

v0.1.2

05 Apr 03:27
Compare
Choose a tag to compare

v0.1.2

  • [2023.04.03] Added support for MPS devices (i.e., M1/M2 Macs). Thanks @brkirch! (Fixes #4)
    • Note: This fix still isn't perfect and may require some extra changes (see issues).
  • [2023.04.04] use_rand now forces itself off if the batch size is odd (meaning the prompted and unprompted images arent in the same batch). This should fix some artifacting without needing to tinker with the settings.

v0.1.1

02 Apr 06:33
Compare
Choose a tag to compare

Skipping v0.1 as it was published externally. Thanks @mkshing for helping set up pypi publishing!

v0.1

2023.03.30

  • Initial release.

2023.03.31

  • Added support for more resolutions than multiples of 16. (Fixes #8)
  • Added support for diffusers (thanks @JunnYu and @ExponentialML)! (Fixes #1)

v0.1.1

2023.04.01

  • Rewrote how the model patching works to address some compatibility issues. (Fixes #9)