This repository presents a full implementation and comparison of two generative models evaluated using the Fréchet Inception Distance (FID) onthe CIFAR‑10 dataset:
-
Flow Matching with Optimal Transport (FlowOT)
leverages optimal transport theory to define straight‑line probability paths and trains a continuous normalizing flow in a simulation‑free manner, dramatically reducing the need for costly ODE solves during training -
Denoising Diffusion Probabilistic Model (DDPM)
A classic 1,000‑step diffusion process with a U‑Net backbone and MSE loss. -
Metrics
FID score, sample quality grids, and inference time benchmarks.
Both are trained for 500 epochs on 50 000 train / 10 000 test CIFAR‑10 images, using only a consumer‑grade GPU (NVIDIA 4060 Ti 16 GB VRAM). Under identical compute budgets, FlowOT achieves FID ≈ 49 vs. DDPM’s ≈ 55, while sampling ∼ 20 % faster.
This repository provides training, eval and visualization scripts. Results can easily be reproduced using the respective model checkpoints.
- Architecture: U‑Net backbone for both FlowOT & DDPM
- DDPM β‑schedule: linear from 0.0001 to 0.02 over 1 000 steps, MSE loss
- Hardware: NVIDIA RTX 4060 Ti (16 GB), 32 GB RAM, Intel i7‑8700K
- Dataset: CIFAR‑10 (32×32 RGB)
- 50 000 training images
- 10 000 test images
- Training: 500 epochs, batch size 128, Adam (β₁ = 0.9, β₂ = 0.999, ϵ = 1e‑8), polynomial LR decay
FlowOT: ~49 FID (vs. DDPM’s ~55); inference time: ~800 ms/image vs. ~1 000 ms/image.
Please see "Paper.pdf" for full details.
.
├── DDPM/ # Implementation of denoising diffusion probabilistic models
├── FlowOT/ # Implementation of flow matching with optimal transport
├── helpers/ # Helper scripts and utilities
├── Paper.pdf # The research paper for the project
├── requirements.txt # Python dependencies
Install dependencies:
pip install -r requirements.txt
Training and evaluation scripts/notebooks are available in respective folders. Checkpoints and training logs can be found in subdirectories of FlowOT and DDPM. Checkpoints are stored with Git LFS.
This project project was created together with my peers mentioned in the paper