Skip to main content
UploadLanguage (EN)Support
BusinessMobileSocial MediaMarketingTechnologyArt & PhotosCareerDesignEducationPresentations & Public SpeakingGovernment & NonprofitHealthcareInternetLawLeadership & ManagementAutomotiveEngineeringSoftwareRecruiting & HRRetailSalesServicesScienceSmall Business & EntrepreneurshipFoodEnvironmentEconomy & FinanceData & AnalyticsInvestor RelationsSportsSpiritualNews & PoliticsTravelSelf ImprovementReal EstateEntertainment & HumorHealth & MedicineDevices & HardwareLifestyle
HL
Uploaded byHyeongmin Lee
PDF, PPTX3,082 views

PR-409: Denoising Diffusion Probabilistic Models

The document discusses denoising diffusion probabilistic models, focusing on their role as generative models similar to variational autoencoders, GANs, and normalizing flows. It highlights techniques for loss simplification and experimental comparisons related to the models. Additionally, it references related literature and resources for further learning on the topic.

Technology◦
Related topics:
Deep Learning•
Download as PDF, PPTX
Denoising Diffusion Probabilistic Models Hyeongmin Lee PR-409 2022.10.30
Intro
Intro  Stable Diffusion
Intro  RAFT [ECCV 2020; PR278]
Intro  Diffusion Model
Generative Models
Generative Models  Generative Models “To know the distribution(manifold) of data” • Variational Autoencoders [PR 010] • GANs [PR 001] • Normalizing Flow https://www.youtube.com/watch?v=TywKpiZJ7P4 𝑧𝑧 𝑥𝑥 High Entropy Low Information Low Entropy High Information
Generative Models  Generative Models
Generative Models  Variational Auto Encoders 𝑧𝑧 � 𝑥𝑥 𝑥𝑥
Generative Models  Generative Adversarial Nets 𝑧𝑧 � 𝑥𝑥 real? fake?
Generative Models  Normalizing Flow Models 𝑧𝑧 � 𝑥𝑥 𝑥𝑥 𝑓𝑓 � 𝑓𝑓 ��� 𝑓𝑓 𝑓𝑓−1 � 𝑓𝑓−1 ��� 𝑓𝑓−1
Generative Models  Diffusion Models 𝑧𝑧 � 𝑥𝑥 𝑥𝑥 𝑞𝑞 � 𝑞𝑞 ��� 𝑞𝑞 𝑝𝑝 � 𝑝𝑝 ��� 𝑝𝑝 Fixed!
Generative Models  Generative Models
Diffusion Probabilistic Models
Diffusion Probabilistic Models  Diffusion?
Diffusion Probabilistic Models  Diffusion Process
Diffusion Probabilistic Models  Diffusion Loss Regularization Reconstruction Regularization Reconstruction 𝛼𝛼𝑡𝑡 = 1 − 𝛽𝛽𝑡𝑡 � 𝛼𝛼𝑡𝑡 = � 𝑠𝑠=1 𝑡𝑡 𝛼𝛼𝑠𝑠 Learning 𝛽𝛽𝑡𝑡
Denoising Diffusion Probabilistic Models (DDPM)
DDPM  Residual Estimation Accurate Image Super-Resolution Using Very Deep Convolutional Networks [CVPR 2016]
DDPM  Loss Simplification
DDPM  Loss Simplification #1 – Deleting Regularization Term “Not to learn, just to fix 𝛽𝛽𝑡𝑡 “
DDPM  Loss Simplification #2 – Not to Learn Variance or
DDPM  Loss Simplification
Experiments
Experiments  Quantitative Comparison
Experiments  Visual Comparison
Experiments  Progressive Generation
Thank You! • Denoising Diffusion Implicit Models [ICLR 2021] • Score-Based Generative Modeling through Stochastic Differential Equations [ICLR 2021] [References] • Youtube Video: [Paper Review] Denoising Diffusion Probabilistic Models (by 김정섭님) • Blog: [논문공부] Denoising Diffusion Probabilistic Models (DDPM) 설명 (by Blue Collar Developer) • Blog: What are Diffusion Models? (by Lil'Log) [Additional]