> ML_ALGORITHM // DENOISING-DIFFUSION-PROBABILISTIC-MODELS-DDPM_v1.0
Denoising Diffusion Probabilistic Models (DDPM)
Generative model class that generates samples by matching a reverse diffusion process that incrementally removes noise from pure Gaussian chaos.
Diffusion & Score-Based Generative Modelsdeep-generativeblack-boxlarge (>100k)
Back to All AlgorithmsComputational Complexity
Training Complexity:O(epochs * batch_size * unet_pass)
Inference Complexity:O(T * unet_pass) where T in [50, 1000]
Hardware Profile
CPU Friendly:No
Requires GPU:Yes
Memory Footprint:high
Interpretability & Data
Interpretability Tier:black-box
Training Data Needs:large (>100k)
Interpretability Assessment
Intermediate diffusion trajectories demonstrate gradual emergence of coarse structural geometry followed by fine details.
Suitable Tasks & Supported Modalities
Suitable Tasks:
image generationaudio synthesismolecular generation
Supported Modalities:
imageaudio3d-pointclouds
Implementing Libraries
Foundational Literature
Deep Unsupervised Learning using Nonequilibrium ThermodynamicsJascha Sohl-Dickstein, Eric A. Weiss (2015) · International Conference on Machine Learning (ICML)
Denoising Diffusion Probabilistic Models (DDPM)Jonathan Ho, Ajay Jain (2020) · Advances in Neural Information Processing Systems (NeurIPS)
Common Pitfalls & Warnings
- Naive inference requires 1,000 forward passes through a heavy U-Net, resulting in multi-second generation latencies without fast samplers