IADB (Iterative Alpha (De)Blending) replaces the rather complex and math-y bits of denoising diffusion probabilistic models with the very simple alpha-blending that all graphics programmers are already familiar with. I love it! Give their paper a read and check out how simple and elegant the derivation is.
Having fun with the iterative alpha (de)blending technique!
- Created pytorch-lightning based trainer to watch pretty graphs in wandb, use exponential moving average to smooth the weights, pre-process training images into LMDB for faster training, and other goodness
- Trained some 64x64 RGB unconditional models
- Config files so can create models with different architectures
- Implemented simple euler and runge-kutta samplers
- Create interpolation animations slerping through noise seeds to see how smooth the learned mappings are
- Created latent-IADB by combining Stable Diffusion VAE with 4-channel IADB to denoise latents
- Trained 512px latent-IADB cat model
This is for fun and learning so don't kick me in the nuts if you try it and something doesn't work.
You can find some models I've trained at https://huggingface.co/pharma/iadb/tree/main and they're available in the droplist inside the notebook. You can't really make anything useful with them (unless you want to fill your drive up with cat pictures).
See trainer.py for full set of parameters. Here's an example:
python trainer.py \
--batch_size 64 \
--image_folder "train/cats" \
--learning_rate 1e-4 \
--lmdb_path "train/cats_lmdb" \
--max_epochs -1 \
--model_config "train/cats512_vae.yaml" \
--out_path "cats/cats512_vae" \
--preview_interval 250 \
--save_ckpt_mins 30 \
--wandb_project iadb_cats_vaeYou need to create a yaml file specifying the number of UNet blocks, channel counts, whether to use VAE etc. Here's example for the 512 cats cats512_vae.yaml