Skip to content

QUICKSTART.md

Latest commit

 

History

History
49 lines (33 loc) · 1.27 KB

QUICKSTART.md

File metadata and controls

49 lines (33 loc) · 1.27 KB

Quick Start Guide

Get started with Flow Matching image generation in 5 minutes.

Installation

pip install -r requirements.txt

Train

Start training on CIFAR-10:

python train_flow_matching.py

The script will:

  • Automatically download CIFAR-10
  • Train a ~50M parameter DiT model
  • Save checkpoints to checkpoints/ every 5000 steps
  • Generate sample images to samples/ every 1000 steps

Generate Images

After training (or use a checkpoint):

# Generate 16 random samples
python generate_flow_matching.py --checkpoint checkpoints/flow_matching_final.pt --num_samples 16

# Generate specific class (e.g., airplane=0, car=1, bird=2, cat=3, deer=4, dog=5, frog=6, horse=7, ship=8, truck=9)
python generate_flow_matching.py --checkpoint checkpoints/flow_matching_final.pt --num_samples 16 --class_label 5

# Generate all classes
python generate_flow_matching.py --checkpoint checkpoints/flow_matching_final.pt --class_grid --samples_per_class 8

Customize

Edit configs/flow_matching_config.py to change:

  • Model size (hidden_size, depth, num_heads)
  • Training settings (batch_size, lr, train_steps)
  • Sampling quality (num_sampling_steps, cfg_scale)

What's Next?

See FLOW_MATCHING_GUIDE.md for detailed documentation.