DDPM (Denoising Diffusion Probabilistic Model)
DDPM is the foundational framework for diffusion models that generates images by progressively denoising from pure noise.
DDPM generates images through progressive denoising – the theoretical foundation behind Stable Diffusion, DALL-E, and all modern image generators.
Explanation
In the forward process, Gaussian noise is gradually added until only noise remains. In the reverse process, a U-Net learns to progressively remove noise. Typical are 1000 forward steps and 20-50 sampling steps with accelerated solvers.
Marketing Relevance
DDPM is the theoretical foundation of all modern image generators – Stable Diffusion, DALL-E, Midjourney build on DDPM principles.
Example
Stable Diffusion 1.5 uses a DDPM-based U-Net in latent space with CLIP text encoder for text-conditioned generation.
Common Pitfalls
Slow sampling (many steps needed). High VRAM requirements. Mode collapse with poor training. Forward/reverse process often confused.
Origin & History
Sohl-Dickstein et al. (2015) introduced diffusion-based generative models. Ho et al. (2020) made them practical with the DDPM paper, surpassing GANs in image quality. Dhariwal & Nichol (2021) showed superiority with "Diffusion Models Beat GANs." DDPM became the basis for Stable Diffusion, DALL-E 2, and Imagen.
Comparisons & Differences
DDPM (Denoising Diffusion Probabilistic Model) vs. GAN
GANs use adversarial training (unstable, mode collapse); DDPM uses stable likelihood-based training with better mode coverage.
DDPM (Denoising Diffusion Probabilistic Model) vs. DDIM
DDPM is stochastic and needs many steps; DDIM is deterministic and can achieve comparable quality with fewer steps (10-20).