Diffusion Models from Scratch

A PyTorch implementation of diffusion models built from scratch, featuring:

Key Features:

🎯 From-Scratch Implementation

• Clean, modular codebase

🚀 Sampling Methods

• DDPM (Denoising Diffusion Probabilistic Models)
• DDIM (Denoising Diffusion Implicit Models)
• (More to come...)

🧠 Network Architectures

• UNet
• DiT (Diffusion Transformer)
• UViT (U-Vision Transformer)

📊 Datasets Support

• MNIST
• CIFAR-10
• (More to come...)

Preliminary

• diffusion process

Result

• result on mnist dataset

• result on cifar-10 dataset

• loss

TODO List

• Introduce cross_attention to fuse image information and classification information
• Implement inference code
• Experiment with different model architectures, e.g., Unet, DiT, UViT, etc.
• Experiment with different noise schedulers, e.g., DDPM, DDIM, DPM-Solver, etc.
• Add evaluation metrics such as FID and CLIP score
• Experiment with different datasets, e.g., MNIST, CIFAR-10, text2image datasets, etc.
• Experiment with different training methods, e.g., LoRA, etc.

Acknowledgements

• Pytorch-diffusion
• Diffusion Transformer
• A ViT Backbone for Diffusion Models
• Tutorial on Diffusion Models for Imaging and Vision
• Probabilistic Diffusion Model概率扩散模型理论与完整PyTorch代码详细解读
• Stable Diffusion 原理介绍与源码分析
• 扩散模型(Diffusion Model)详解：直观理解、数学原理、PyTorch 实现