Flow Diffusion Lab

Research project exploring:

• Diffusion models
• Flow matching
• Transformer-based generative models

Goal

Compare diffusion vs flow matching for image generation.

Dataset

CIFAR10 (initial)

Models

• Diffusion Transformer
• Flow matching vector field model

Setup

pip install -r requirements.txt

Training

Trains the flow-matching DiT on CIFAR10 and saves a checkpoint to checkpoints/:

python train.py --epochs 10

Useful flags for a quick run: --max-steps 200 --subset-size 512 --checkpoint-name my_run.

Comparing checkpoints

Launch a Gradio app to pick two saved checkpoints and compare their generations, starting from the same noise seed:

python app.py