Diffusion Models
Diffusion Models 1
Implement the mathematics that surrounds a diffusion model's neural network: noise schedules, the closed-form forward process, SNR-based loss weighting, the DDPM posterior, DDIM sampling, classifier-free guidance and latent scaling. Every exercise takes the network's prediction as a given array, so you write exactly the code that lives outside the U-Net in a real sampler.
50
problems
5
topics
15
free
01
DDPM and DDIM Sampling
10 problems- 1Predict x0 from a Noise EstimateFREEEasy
- 2DDIM Deterministic Step (eta = 0)PROEasy
- 3One DDPM Ancestral Sampling StepFREEMedium
- 4DDIM Timestep SubsequencePROMedium
- 5Full DDPM Ancestral StepPROMedium
- 6DDIM with Stochasticity (general eta)PROMedium
- 7Uniform Stride Timestep SubsequencePROMedium
- 8Quadratic Stride Timestep SubsequencePROMedium
- 9The DDIM Update RuleFREEHard
- 10Ancestral Sampling Trajectory NormPROHard
02
Forward Process and Noise Schedules
10 problems- 11Linear Beta Schedule and Alpha-BarFREEEasy
- 12Closed-Form Forward Sample q(x_t | x_0)PROEasy
- 13Quadratic Beta SchedulePROEasy
- 14Alpha-Bar to Beta InversionPROEasy
- 15Cosine Noise ScheduleFREEMedium
- 16Sigmoid Beta SchedulePROMedium
- 17Forward-Diffuse a Batch to Timestep tPROMedium
- 18Terminal SNR and the Zero-SNR FixPROMedium
- 19Noise Level at a Target SNRPROMedium
- 20Karras Sigma ScheduleFREEHard
03
Guidance and Latent Diffusion
10 problemsSign up free to see all 50 problems
Create a free account to see the full Diffusion Models 1 list, track your progress, and save problems to lists.
Sign up free