Image Quality and Sample Diversity
What you'll learn: How DDPMs achieve both high-quality images and rich diversity by leveraging their gradual denoising process.
The Dual Achievement of DDPMs
One of the most remarkable features of Denoising Diffusion Probabilistic Models is their ability to generate images that are both visually sharp and varied. This is a hard balance to strike—GANs often suffer from mode collapse (low diversity), while VAEs can produce diverse but blurry samples.
Why DDPMs Excel at Both
High Fidelity Through Gradual Denoising: Think of painting a detailed portrait. If you try to paint it all at once, you might make mistakes. But if you refine it gradually—first sketching outlines, then adding colors, then fine details—you achieve better quality. DDPMs work similarly: they remove noise step-by-step across many timesteps (often 1000+), allowing the denoising network to make small, careful corrections at each stage. This incremental refinement produces crisp, detailed images.
Diversity Through Stochastic Sampling: During the reverse process (ancestral sampling), DDPMs add small amounts of random noise at each timestep before denoising. This stochasticity ensures that starting from pure Gaussian noise leads to different final images each time. Unlike a deterministic decoder, every sampling run explores different paths through the denoising chain, naturally producing diverse outputs without sacrificing quality.
The Balance in Action
The variance schedule controls this trade-off: smaller beta values mean gentler noise addition, allowing finer control during denoising (better quality), while the stochastic sampling preserves the richness of the learned distribution (diversity).
Key Takeaway: DDPMs achieve high image quality through gradual, multi-step denoising and maintain sample diversity through stochastic ancestral sampling—solving a key weakness of both GANs and VAEs.