# explainingai-code/dit-pytorch architecture

> Architecture diagram of the GitHub repository explainingai-code/dit-pytorch, generated by GitDiagram from its file tree, README and a sample of source files (last updated 2026-10-03). It is an AI-made overview: check details against the source.

- Interactive diagram: https://gitdiagram.com/explainingai-code/dit-pytorch
- Repository: https://github.com/explainingai-code/dit-pytorch

## Overview

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: [`tools/sample_vae_dit.py`](https://github.com/explainingai-code/dit-pytorch/blob/main/tools/sample_vae_dit.py) (DiT sampling), [`dataset/celeb_dataset.py`](https://github.com/explainingai-code/dit-pytorch/blob/main/dataset/celeb_dataset.py) (Image dataset), [`tools/train_vae.py`](https://github.com/explainingai-code/dit-pytorch/blob/main/tools/train_vae.py) (Train VAE), [`tools/train_vae_dit.py`](https://github.com/explainingai-code/dit-pytorch/blob/main/tools/train_vae_dit.py) (Train DiT)

## Diagram (Mermaid)

```mermaid
flowchart TD

subgraph group_data["Data and latents"]
  node_images[("CelebHQ images")]
  node_dataset["Image dataset<br/>[celeb_dataset.py]"]
  node_latents[("Saved VAE latents<br/>[diffusion_utils.py]")]
end

subgraph group_vae["VAE training"]
  node_vae_train["Train VAE<br/>[train_vae.py]"]
  node_perceptual["LPIPS metric<br/>[lpips.py]"]
  node_discriminator["Discriminator<br/>[discriminator.py]"]
  node_vae_infer["VAE inference<br/>[infer_vae.py]"]
end

subgraph group_dit["DiT diffusion"]
  node_dit_train["Train DiT<br/>[train_vae_dit.py]"]
  node_scheduler["Noise scheduler"]
  node_sampling["DiT sampling<br/>[sample_vae_dit.py]"]
  node_generated[("Generated images")]
end

subgraph group_architecture["Model architecture"]
  node_vae["VAE encoder/decoder<br/>[vae.py]"]
  node_vae_blocks["VAE blocks<br/>[blocks.py]"]
  node_dit["DiT transformer<br/>[transformer.py]"]
  node_patch["Patch embedding<br/>[patch_embed.py]"]
  node_layer["Transformer layer"]
  node_attention["Attention<br/>[attention.py]"]
end

node_user(("Researcher"))
node_config["CelebHQ config<br/>[celebhq.yaml]"]

node_user -->|"configures runs"| node_config
node_user -->|"starts training"| node_vae_train
node_user -->|"runs inference"| node_vae_infer
node_user -->|"starts training"| node_dit_train
node_user -->|"requests samples"| node_sampling
node_images -->|"provides images"| node_dataset
node_dataset -->|"supplies batches"| node_vae_train
node_vae_train -->|"trains"| node_vae
node_vae_train -->|"evaluates perceptual loss"| node_perceptual
node_vae_train -->|"trains adversarially"| node_discriminator
node_vae -->|"uses blocks"| node_vae_blocks
node_vae_infer -->|"loads images"| node_dataset
node_vae_infer -->|"encodes and decodes"| node_vae
node_vae_infer -->|"saves latents"| node_latents
node_dataset -->|"loads cached latents"| node_latents
node_dataset -->|"supplies training data"| node_dit_train
node_dit_train -->|"encodes when needed"| node_vae
node_dit_train -->|"trains noise predictor"| node_dit
node_dit_train -->|"adds training noise"| node_scheduler
node_dit -->|"patchifies input"| node_patch
node_dit -->|"applies layers"| node_layer
node_layer -->|"uses attention"| node_attention
node_sampling -->|"predicts noise"| node_dit
node_sampling -->|"steps backward"| node_scheduler
node_sampling -->|"decodes final latent"| node_vae
node_sampling -->|"writes image samples"| node_generated

click node_config "https://github.com/explainingai-code/dit-pytorch/blob/main/config/celebhq.yaml"
click node_dataset "https://github.com/explainingai-code/dit-pytorch/blob/main/dataset/celeb_dataset.py"
click node_latents "https://github.com/explainingai-code/dit-pytorch/blob/main/utils/diffusion_utils.py"
click node_vae_train "https://github.com/explainingai-code/dit-pytorch/blob/main/tools/train_vae.py"
click node_vae "https://github.com/explainingai-code/dit-pytorch/blob/main/model/vae.py"
click node_vae_blocks "https://github.com/explainingai-code/dit-pytorch/blob/main/model/blocks.py"
click node_perceptual "https://github.com/explainingai-code/dit-pytorch/blob/main/model/lpips.py"
click node_discriminator "https://github.com/explainingai-code/dit-pytorch/blob/main/model/discriminator.py"
click node_vae_infer "https://github.com/explainingai-code/dit-pytorch/blob/main/tools/infer_vae.py"
click node_dit_train "https://github.com/explainingai-code/dit-pytorch/blob/main/tools/train_vae_dit.py"
click node_scheduler "https://github.com/explainingai-code/dit-pytorch/blob/main/scheduler/linear_scheduler.py"
click node_dit "https://github.com/explainingai-code/dit-pytorch/blob/main/model/transformer.py"
click node_patch "https://github.com/explainingai-code/dit-pytorch/blob/main/model/patch_embed.py"
click node_layer "https://github.com/explainingai-code/dit-pytorch/blob/main/model/transformer_layer.py"
click node_attention "https://github.com/explainingai-code/dit-pytorch/blob/main/model/attention.py"
click node_sampling "https://github.com/explainingai-code/dit-pytorch/blob/main/tools/sample_vae_dit.py"

classDef toneNeutral fill:#f8fafc,stroke:#334155,stroke-width:1.5px,color:#0f172a
classDef toneBlue fill:#dbeafe,stroke:#2563eb,stroke-width:1.5px,color:#172554
classDef toneAmber fill:#fef3c7,stroke:#d97706,stroke-width:1.5px,color:#78350f
classDef toneMint fill:#dcfce7,stroke:#16a34a,stroke-width:1.5px,color:#14532d
classDef toneRose fill:#ffe4e6,stroke:#e11d48,stroke-width:1.5px,color:#881337
classDef toneIndigo fill:#e0e7ff,stroke:#4f46e5,stroke-width:1.5px,color:#312e81
classDef toneTeal fill:#ccfbf1,stroke:#0f766e,stroke-width:1.5px,color:#134e4a
class node_images,node_dataset,node_latents toneBlue
class node_vae_train,node_perceptual,node_discriminator,node_vae_infer toneAmber
class node_dit_train,node_scheduler,node_sampling,node_generated toneMint
class node_vae,node_vae_blocks,node_dit,node_patch,node_layer,node_attention toneRose
class node_user toneIndigo
class node_config toneTeal
```

## Components

### Data and latents

- **CelebHQ images**
- [Image dataset](https://github.com/explainingai-code/dit-pytorch/blob/main/dataset/celeb_dataset.py): `dataset/celeb_dataset.py`
- [Saved VAE latents](https://github.com/explainingai-code/dit-pytorch/blob/main/utils/diffusion_utils.py): `utils/diffusion_utils.py`

### VAE training

- [Train VAE](https://github.com/explainingai-code/dit-pytorch/blob/main/tools/train_vae.py): `tools/train_vae.py`
- [LPIPS metric](https://github.com/explainingai-code/dit-pytorch/blob/main/model/lpips.py): `model/lpips.py`
- [Discriminator](https://github.com/explainingai-code/dit-pytorch/blob/main/model/discriminator.py): `model/discriminator.py`
- [VAE inference](https://github.com/explainingai-code/dit-pytorch/blob/main/tools/infer_vae.py): `tools/infer_vae.py`

### DiT diffusion

- [Train DiT](https://github.com/explainingai-code/dit-pytorch/blob/main/tools/train_vae_dit.py): `tools/train_vae_dit.py`
- [Noise scheduler](https://github.com/explainingai-code/dit-pytorch/blob/main/scheduler/linear_scheduler.py): `scheduler/linear_scheduler.py`
- [DiT sampling](https://github.com/explainingai-code/dit-pytorch/blob/main/tools/sample_vae_dit.py): `tools/sample_vae_dit.py`
- **Generated images**

### Model architecture

- [VAE encoder/decoder](https://github.com/explainingai-code/dit-pytorch/blob/main/model/vae.py): `model/vae.py`
- [VAE blocks](https://github.com/explainingai-code/dit-pytorch/blob/main/model/blocks.py): `model/blocks.py`
- [DiT transformer](https://github.com/explainingai-code/dit-pytorch/blob/main/model/transformer.py): `model/transformer.py`
- [Patch embedding](https://github.com/explainingai-code/dit-pytorch/blob/main/model/patch_embed.py): `model/patch_embed.py`
- [Transformer layer](https://github.com/explainingai-code/dit-pytorch/blob/main/model/transformer_layer.py): `model/transformer_layer.py`
- [Attention](https://github.com/explainingai-code/dit-pytorch/blob/main/model/attention.py): `model/attention.py`

### Other components

- **Researcher**
- [CelebHQ config](https://github.com/explainingai-code/dit-pytorch/blob/main/config/celebhq.yaml): `config/celebhq.yaml`

## Connections

- Researcher → CelebHQ config: configures runs (evidence: `README.md`)
- Researcher → Train VAE: starts training (evidence: `README.md`)
- Researcher → VAE inference: runs inference (evidence: `README.md`)
- Researcher → Train DiT: starts training (evidence: `README.md`)
- Researcher → DiT sampling: requests samples (evidence: `README.md`)
- CelebHQ images → Image dataset: provides images (evidence: `dataset/celeb_dataset.py`)
- Image dataset → Train VAE: supplies batches (evidence: `tools/train_vae.py`)
- Train VAE → VAE encoder/decoder: trains (evidence: `tools/train_vae.py`)
- Train VAE → LPIPS metric: evaluates perceptual loss (evidence: `tools/train_vae.py`)
- Train VAE → Discriminator: trains adversarially (evidence: `tools/train_vae.py`)
- VAE encoder/decoder → VAE blocks: uses blocks (evidence: `model/vae.py`)
- VAE inference → Image dataset: loads images (evidence: `tools/infer_vae.py`)
- VAE inference → VAE encoder/decoder: encodes and decodes (evidence: `tools/infer_vae.py`)
- VAE inference → Saved VAE latents: saves latents (evidence: `README.md`)
- Image dataset → Saved VAE latents: loads cached latents (evidence: `dataset/celeb_dataset.py`)
- Image dataset → Train DiT: supplies training data (evidence: `tools/train_vae_dit.py`)
- Train DiT → VAE encoder/decoder: encodes when needed (evidence: `tools/train_vae_dit.py`)
- Train DiT → DiT transformer: trains noise predictor (evidence: `tools/train_vae_dit.py`)
- Train DiT → Noise scheduler: adds training noise (evidence: `tools/train_vae_dit.py`)
- DiT transformer → Patch embedding: patchifies input (evidence: `model/transformer.py`)
- DiT transformer → Transformer layer: applies layers (evidence: `model/transformer.py`)
- Transformer layer → Attention: uses attention (evidence: `model/transformer_layer.py`)
- DiT sampling → DiT transformer: predicts noise (evidence: `tools/sample_vae_dit.py`)
- DiT sampling → Noise scheduler: steps backward (evidence: `tools/sample_vae_dit.py`)
- DiT sampling → VAE encoder/decoder: decodes final latent (evidence: `tools/sample_vae_dit.py`)
- DiT sampling → Generated images: writes image samples (evidence: `README.md`)
