ImageReconstruction

Diffusion and Autoencoder Image Models

An older CS 472 (unsupervised learning) course project. It trains DDPM diffusion models on MNIST and CIFAR-10, with autoencoder, denoising autoencoder, and variational autoencoder baselines on MNIST. Code: github.com/kalebcoleman/ImageReconstruction.

Diffusion

Both models are trained from scratch and sampled with 50 DDIM steps. The settings below are the ones in the repository’s configs.

	MNIST	CIFAR-10
Image size	28x28 grayscale	32x32 RGB
Network	small U-Net	ADM-style U-Net with attention at 8x8
Timesteps and schedule	500, linear	1,000, cosine
Prediction target	noise (epsilon)	v-prediction
Conditioning	none	class labels with classifier-free guidance (scale 3.0)
Training length	50 epochs	150 epochs, EMA 0.999

The grids below are generated samples, enlarged 4x with nearest-neighbor scaling so the browser does not blur them. The samples are shown as generated; no quantitative sample-quality metric such as FID is reported here.

MNIST samples

A 10 by 10 grid of handwritten digits generated by the MNIST diffusion model

CIFAR-10 samples

Each column is one class.

A 10 by 10 grid of small color images generated by the CIFAR-10 diffusion model, one class per column: airplanes, cars, birds, cats, deer, dogs, frogs, horses, ships and trucks

Autoencoder baselines

Three models reconstruct MNIST digits through a 16-dimensional latent space: a regular autoencoder (AE), a denoising autoencoder (DAE) trained on inputs with noise level 0.2, and a variational autoencoder (VAE).

Bar charts of mean squared error, PSNR and SSIM for the AE, DAE and VAE on MNIST at latent dimension 16

Autoencoder reconstructions:

MNIST digits and their autoencoder reconstructions at latent dimension 16

Denoising autoencoder, reconstructing from noisy inputs:

Noisy MNIST digits and their denoising autoencoder reconstructions at latent dimension 16

Samples drawn from the VAE’s latent distribution:

MNIST-like digits sampled from the variational autoencoder

Scope

This is coursework on small benchmark datasets, kept as a record. The course project also planned a PCA comparison done by a teammate. That part is not in this repository, so no PCA-versus-autoencoder conclusions are drawn here.

Reproduce

pip install -r requirements.txt

# Diffusion
python train.py --config configs/diffusion/mnist.yaml
python train.py --config configs/diffusion/cifar10.yaml

# Autoencoder baselines
python train.py --models ae dae vae --dataset mnist --latent-dim 16

This site is open source. Improve this page.