Build from first principles

Generate Images with Diffusion and Rectified Flow

Add noise to data, train a network to remove it, and turn fresh noise into new samples with DDPM. Then make the paths straight with rectified flow and find out what happens when you only have a few steps.

What is a diffusion model?See how destroying data with noise can teach a network to create it

1Turn a Noise Schedule into Signal LevelsHow fast the signal fades

2Jump Straight to Noise Step tNoising in one jump

3Compute the DDPM Training LossLearning to predict the noise

4Take One Reverse Diffusion StepOne step back toward data

5Sample a Rectified Flow with Euler StepsFollowing a velocity field

6Sample a DDPM with Only K StepsSkipping steps

From Noise to ImagesSee how your functions fit together before sampling

Sample a Spiral and Clothing ImagesTrain DDPM and rectified flow, sample with 1000 steps, then with only a few