v0.1.2

What's Changed

• NoOp Model by @Landanjs in #139
• Script to pre-compute CLIP and T5 by @Landanjs in #144
• Add option to shift noise schedules when changing resolution by @coryMosaicML in #153
• Expose option to set per-stream weighting in image and image_caption datasets by @coryMosaicML in #156
• HF image generation that integrates with Cory's earlier script by @rishab-partha in #158
• MMDiT implementation and text-to-image training with rectified flows by @coryMosaicML in #155
• Add option to use predefined aspect ratio buckets in the cropping transform by @coryMosaicML in #157
• Add latent logger for T5-XXL text encoder by @rishab-partha in #154
• Pass loggers to Trainer in eval by @jazcollins in #166
• Simple LoRA Finetuning (WIP) by @rishab-partha in #164
• Add option to change start and end SNR in SD2/SDXL configs by @coryMosaicML in #165
• Small bug fixes to bulk image generation by @coryMosaicML in #167
• Add dataset for running with precomputed latents from multiple captions by @coryMosaicML in #161
• Small bug fixes for running models without tokenizers by @coryMosaicML in #168

New Contributors

• @rishab-partha made their first contribution in #158

Full Changelog: v0.1.1...v0.1.2

v0.1.1

Minor bug fix related to max_pad_tokens at generate time. Other Noise Schedule related features and options.

What's Changed

• Optional quasirandom timesteps, zero terminal SNR, cosine schedule for SD models by @coryMosaicML in #138
• Add HF hub dependency by @coryMosaicML in #142
• Add link to CommonCanvas model weights by @Skylion007 in #143
• Fix autoencoder load by @RR4787 in #141
• Add option to use karras sigmas for SDXL style models by @coryMosaicML in #146
• Fix bug in stable diffusion when mask_pad_tokens is false by @coryMosaicML in #147
• Only use a text encoder mask in SD model forward if mask_pad_tokens is false by @coryMosaicML in #149

Full Changelog: v0.1.0...v0.1.1

v0.1.0

What's Changed

• Allow masking padding tokens in cross attention layers by @jazcollins in #94
• Fix typo in pyproject.toml by @eltociear in #92
• Autoencoder implementation and training by @coryMosaicML in #79
• Hotfix missing lpips requirement by @Skylion007 in #98
• Fixes for doing inference with masked padding by @coryMosaicML in #99
• Add script for running gradio demo from a local checkpoint by @coryMosaicML in #100
• Better StreamingDataset defaults while preserving old shuffle settings by @snarayan21 in #95
• Remove rounding in aspect ratio bucketing transform by @Landanjs in #111
• Add sample SDXL yamls and update README by @jazcollins in #112
• LogDiffusionImages Features and Refactors by @Landanjs in #104
• Bump gradio demo version by @coryMosaicML in #114
• Make custom autoencoders work with SD2 and SDXL models. by @coryMosaicML in #102
• Add algorithm to control randomness over different eval times by @coryMosaicML in #115
• Add ruff linter by @Skylion007 in #122
• Add image only dataset + script to add captions generated by LLaVA to a streaming dataset by @coryMosaicML in #118
• Code cleanup by @coryMosaicML in #120
• Update to latest transformers, diffusers, and other packages. by @coryMosaicML in #125
• Landan/text encoder refactor by @Landanjs in #124
• Add option to specify image output key in image dataloader factory by @coryMosaicML in #129
• Add explicit per block fsdp wrapping for SDXL by @coryMosaicML in #127
• make local paths optional by @A-Jacobson in #128
• fixed masked padding bug. by @A-Jacobson in #130
• arbitrary aspect ratio buckets by @Landanjs in #126
• Add option to set per-channel mean, std. dev. of the autoencoder latents when training the UNet by @coryMosaicML in #132
• Test PR by @Landanjs in #134
• Arbitrary aspect ratio bucket boundaries by @Landanjs in #133
• Bug to enable fp16 by @RR4787 in #136
• Only download CLIP on rank 0 when doing eval by @coryMosaicML in #135

New Contributors

• @eltociear made their first contribution in #92
• @snarayan21 made their first contribution in #95
• @RR4787 made their first contribution in #136

Full Changelog: v0.0.1...v0.1.0

v0.0.1

Add callback to catch NaNs in the train loss (#97)