Release v0.1.0

Better Pivotal Tuning, CLIP evaluation, SVD distillation!

Many tricks / features went into this release :

1. Extended Latent textual inversion
2. Pivotal tuning with stage-learning rate schedule
3. Norm-prior Bayesian learning
4. Mask conditioned score estimation (+ in-pipeline face detector)
5. Sub-1 denoising
6. CLIP score (text alignment score, image alignment score) evaluater
7. xformer support (not in lorpt CLI yet)
8. wandb logging
9. Better Caption dataset
10. Distilling fully trained model into LoRA : SVD distillation CLI

What's Changed

• fix(missing-var): h_flip by @oscarnevarezleal in #86
• BetterPTI: Support pivotal tuning with multiword latent by @cloneofsimo in #91
• Other Acceleration tricks by @cloneofsimo in #93
• fix : save lora 16 precision by @cloneofsimo in #97
• feat : svd distillation with CLI by @cloneofsimo in #98
• Add xformers to training scripts by @hafriedlander in #103
• Make safetensors properly optional, and support storing TI by @hafriedlander in #101
• Feat/inversion pp by @cloneofsimo in #104
• image transforms by @ethansmith2000 in #115
• Feat/allsave by @cloneofsimo in #119
• v 0.1.0 by @cloneofsimo in #88

New Contributors

• @timh made their first contribution in #11
• @AK391 made their first contribution in #20
• @2kpr made their first contribution in #25
• @DavidePaglieri made their first contribution in #39
• @hdeezy made their first contribution in #48
• @hafriedlander made their first contribution in #73
• @milyiyo made their first contribution in #75
• @laksjdjf made their first contribution in #76
• @oscarnevarezleal made their first contribution in #86
• @ethansmith2000 made their first contribution in #115

Full Changelog: https://github.com/cloneofsimo/lora/commits/v0.1.0-alpha