How explainingai-code/dit-pytorch works
This PyTorch repository trains a VAE for CelebHQ, then trains a configurable Diffusion Transformer on the VAE latent space and samples images by reversing the diffusion process and decoding the final latent. The graph preserves the documented VAE inference/reconstruction and latent-saving branch, as well as the optional precomputed-latent path. Edges reflect README statements or shown imports and calls; unsampled implementation details are not inferred.
Main entry points: sample_vae_dit.py DiT sampling · celeb_dataset.py Image dataset · train_vae.py Train VAE · train_vae_dit.py Train DiT
Components 19
Data and latents
- CelebHQ images
- Image dataset
dataset/celeb_dataset.py - Saved VAE latents
utils/diffusion_utils.py
VAE training
- Train VAE
tools/train_vae.py - LPIPS metric
model/lpips.py - Discriminator
model/discriminator.py - VAE inference
tools/infer_vae.py
DiT diffusion
- Train DiT
tools/train_vae_dit.py - Noise scheduler
scheduler/linear_scheduler.py - DiT sampling
tools/sample_vae_dit.py - Generated images
Model architecture
- VAE encoder/decoder
model/vae.py - VAE blocks
model/blocks.py - DiT transformer
model/transformer.py - Patch embedding
model/patch_embed.py - Transformer layer
model/transformer_layer.py - Attention
model/attention.py
Other components
- Researcher
- CelebHQ config
config/celebhq.yaml
Connections 26
- Researcher → CelebHQ config: configures runs
README.md - Researcher → Train VAE: starts training
README.md - Researcher → VAE inference: runs inference
README.md - Researcher → Train DiT: starts training
README.md - Researcher → DiT sampling: requests samples
README.md - CelebHQ images → Image dataset: provides images
dataset/celeb_dataset.py - Image dataset → Train VAE: supplies batches
tools/train_vae.py - Train VAE → VAE encoder/decoder: trains
tools/train_vae.py - Train VAE → LPIPS metric: evaluates perceptual loss
tools/train_vae.py - Train VAE → Discriminator: trains adversarially
tools/train_vae.py - VAE encoder/decoder → VAE blocks: uses blocks
model/vae.py - VAE inference → Image dataset: loads images
tools/infer_vae.py - VAE inference → VAE encoder/decoder: encodes and decodes
tools/infer_vae.py - VAE inference → Saved VAE latents: saves latents
README.md - Image dataset → Saved VAE latents: loads cached latents
dataset/celeb_dataset.py - Image dataset → Train DiT: supplies training data
tools/train_vae_dit.py - Train DiT → VAE encoder/decoder: encodes when needed
tools/train_vae_dit.py - Train DiT → DiT transformer: trains noise predictor
tools/train_vae_dit.py - Train DiT → Noise scheduler: adds training noise
tools/train_vae_dit.py - DiT transformer → Patch embedding: patchifies input
model/transformer.py - DiT transformer → Transformer layer: applies layers
model/transformer.py - Transformer layer → Attention: uses attention
model/transformer_layer.py - DiT sampling → DiT transformer: predicts noise
tools/sample_vae_dit.py - DiT sampling → Noise scheduler: steps backward
tools/sample_vae_dit.py - DiT sampling → VAE encoder/decoder: decodes final latent
tools/sample_vae_dit.py - DiT sampling → Generated images: writes image samples
README.md