Brief
Write a CIFAR-10 U-Net (an encoder-decoder CNN with skip connections between matching resolutions) with residual blocks, attention at 16×16, and timestep embeddings (a vector that tells the network how noisy its input is). Train it twice: once as DDPM with ε-prediction (the network predicts the noise that was added to an image) and once with conditional flow matching (the network predicts the velocity that moves a noise sample in a straight line toward a data sample). Implement DDPM, DDIM (a deterministic sampler that can skip steps) and Euler/Heun ODE samplers (first- and second-order step rules for following the learned velocity). Then compute FID-50k with pytorch-fid at NFE 10, 50 and 250.
VICTORY CONDITIONS
- best FID ≤ 15 within 12 GPU-hours; a clear equal-budget comparison of DDPM vs flow matching at NFE 10/50/250 (FID-50k (Fréchet Inception Distance over 50k generated images; lower means more realistic and varied) vs number of function evaluations (NFE, the number of network calls one sample takes), at a fixed 12 GPU-hour training budget).
- A public repository with a README that shows your result and how to reproduce it.
Terms in plain words
- DDPM
- Denoising diffusion: add noise to images step by step, train a network to remove it, then generate by denoising pure noise.
- Flow matching
- Train a network to predict the direction from a noise sample to a data sample, then generate by following that direction with an ODE solver.
- DDIM
- A diffusion sampler with no randomness that can take big steps, so it needs far fewer network calls.
- NFE
- Number of function evaluations: how many times the network runs to make one image.
- FID
- A distance between the feature statistics of real and generated images; lower is better.
- U-Net
- A CNN that downsamples then upsamples, with shortcuts that copy details from the down path to the up path.
Why it rates ★★★☆☆
A controlled equal-compute comparison is more useful than another 'my diffusion model' demo, and generative modelling is a hiring area.