Skip to content

README.md

Latest commit

 

History

History
32 lines (23 loc) · 1.4 KB

README.md

File metadata and controls

32 lines (23 loc) · 1.4 KB
Raw

Diffusion Models

This directory contains standalone diffusion-family implementations in the same repo style used by the neighboring computer-vision families: serially numbered folders, no shared runtime helpers, local dataset routing, local checkpoints, and TensorFlow/Keras-first multi-GPU training entrypoints.

Folder Order

  1. 1_DiffusionProbabilisticModels
  2. 2_ImprovedDDPM
  3. 3_DDIM
  4. 4_GLIDE
  5. 5_Imagen
  6. 6_DALLE2
  7. 7_StableDiffusion
  8. 8_ControlNet
  9. 9_ConsistencyModels

Common Structure

Each folder is standalone and includes:

  • config.py for dataset roots and hyperparameters
  • dataset.py for fully local dataset loading and preprocessing
  • model.py for the architecture and training logic of that folder only
  • train_and_test.py for CLI, multi-GPU setup, checkpoints, resume, and sampling
  • run.sh for repeated local runs
  • README.md for scope and dataset notes

Practical Scope

These folders are designed to be recognizable implementations rather than full paper-scale reproductions. The code keeps the repo's from-scratch style, conservative defaults, local caption parsing, and practical multi-GPU support.

Text-conditioned folders default to local COCO captions and support Flickr30k as a lighter fallback. ControlNet uses on-the-fly edge conditioning, and latent-diffusion folders keep their VAE and latent UNet local to each folder rather than sharing code.