v0.2.2 - astronaut edition

SDXL v_prediction, (under)trained via SimpleTuner:

What's Changed

• Resolve an issue with over-sampling of images, increasing randomness
• Resolve post-training validation error
  • Both by @bghira in #50

Full Changelog: v0.2.1...v0.2.2

v0.2.1

What's Changed

• Next-gen data loader and aspect bucket infrastructure by @bghira in #37
• Diffusers upstream fix for warmup step count
• Improved accuracy in bucket sampling (more random)

Full Changelog: v0.2.0...v0.2.1

v0.2.0 - Experimental, beware!

What's Changed

• Next-gen data loader and aspect bucket infrastructure that is incompatible with previous releases
  • Split the aspect bucket sampler and dataset code out so that it is more readable, more maintainable.
  • VAE cache now uses image filenames instead of SHA hashes. This speeds up creation, and makes it easier to remove the cache entries at some future point.
• More efficient aspect bucketing, lower system memory use, higher CPU utilization, better saturation of disk I/O
• Aspect buckets are now rounded down to two decimals
• Prompt handler class for managing captions
• Working SDXL trainer, thanks to image normalization fix

Full Changelog: v0.1.10...v0.2.0

v0.1.11 - Min-SNR weighted loss, and dropout fixes

What's Changed

• Luminance tracking for WandB by @bghira in #31
  • Track training luminance vs validation luminance
• Dropout: add SDXL caption dropout compatibility by @bghira in #32
• Logging luminance for training data
• Input perturbation for SDXL (Experimental)
• Caption dropout fix for SDXL
• and, Save 'seen' state for SDXL aspect bucketing by @bghira in #33
• Min-SNR weighted loss, helps converge more quickly. (Experimental)

Full Changelog: v0.1.8...v0.1.11

v0.1.9

What's Changed

• Import SimpleTuner prompt validation library via --validation_prompt_library by @bghira in #29
• New dataloader arguments for deleting problematic images by @bghira in #30
• Arguments: add terminal SNR parameters for tweaking, rather than being baked-in
• Make terminal SNR opt-in

Full Changelog: v0.1.8...v0.1.9

v0.1.8 - Even smaller datasets!

What's Changed

• Fixes for some small-dataset training issues by @bghira in #24
• Moving things around by @bghira in #25
• Validation image settings now exist by @bghira in #26

Full Changelog: v0.1.7...v0.1.8

v0.1.7 - Now comes in a smaller size

What's Changed

• Release by @bghira in #18
• allow designating max number of epochs, and fix Triton / python dependencies by @bghira in #19
• Update ubuntu installation script for NVIDIA dependencies by @bghira in #20
• Update LAION dataset script for camera model detection by @bghira in #21
• Release by @bghira in #22
• Aspect buckets with small datasets and a single bucket should no longer error by @bghira in #23

Full Changelog: v0.1.6...v0.1.7

v0.1.6

What's Changed

• Merge parameter parsing between SDXL and SD 2.1
• Add prediction_type parameter by @bghira in #13
• aspect bucketing fix and noise scheduler change by @bghira in #14
• Release by @bghira in #15
• Revert noise scheduler fixes. by @bghira in #16
• Single bucket aspecting failure by @bghira in #17

Full Changelog: v0.1.4...v0.1.6

v0.1.5 - buckets for aspects, really puts things in perspective

What's Changed

• minor updates by @bghira in #3
• VAE Cache improvements by @bghira in #4

Full Changelog: v0.1.0...v0.1.5

v0.1.4 - fix the things we thought we fixed before

What's Changed

• updates and fixes by @bghira in #12

Full Changelog: v0.1.3...v0.1.4