DDPM(Denoising Diffusion Probabilistic Models)과 DDIM(Denoising Diffusion Implicit Modles) 분석

Diffusion Model의 시초인 Diffusion Probabilistic Models부터 Score-based Generative Model(NCSN), Denoising Diffusion Probabilistic Models(DDPM) 그리고 Denoising Diffusion Implicit Models(DDIM)까지 정리하는 시리즈의 두 번째 글에서는 DDPM과 DDIM에 관해 리뷰해 볼 것이다.

 

 

1. DDPM(Denoising Diffusion Probabilistic Models)

 

(1) Key Point

1) Jonathan Ho et al., Denoising Diffusion Probabilistic Models.” arXiv:2006.11239 (2019)

 

 

(2) Review

 

DDPM에 관해 수식적으로 자세히 정리한 글은 아래 링크를 참조.

https://glanceyes.tistory.com/entry/Generative-Model%EA%B3%BC-Diffusion-Model-%EA%B7%B8%EB%A6%AC%EA%B3%A0-Denoising-Diffusion-Probabilistic-Model

Generative Model과 Diffusion Model, 그리고 Denoising Diffusion Probabilistic Model

 

 

1) Jonathan Ho et al., “Denoising Diffusion Probabilistic Models.” arXiv:2006.11239 (2019)

 

 

1) Jonathan Ho et al., “Denoising Diffusion Probabilistic Models.” arXiv:2006.11239 (2019)

 

 

(3) Implementation

 

DDPM을 from scratch로 간단하게 구현한 코드는 아래 링크를 참조.

https://github.com/Glanceyes/ML-Paper-Review/blob/main/ComputerVision/Diffusion/DDPM/DDPM.ipynb

GitHub - Glanceyes/ML-Paper-Review: Implementation of ML&DL models in machine learning that I have studied and written source co

 

 

1) Jonathan Ho et al., “Denoising Diffusion Probabilistic Models.” arXiv:2006.11239 (2019)

 

 

1) Jonathan Ho et al., “Denoising Diffusion Probabilistic Models.” arXiv:2006.11239 (2019)

 

 

 

1) Jonathan Ho et al., “Denoising Diffusion Probabilistic Models.” arXiv:2006.11239 (2019)

 

 

(4) Problem

1) Jonathan Ho et al., “Denoising Diffusion Probabilistic Models.” arXiv:2006.11239 (2019)

 

 

1) Jonathan Ho et al., “Denoising Diffusion Probabilistic Models.” arXiv:2006.11239 (2019)

 

 

 

2. Denoising Diffusion Implicit Models

 

(1) Key Point

2) Jiaming Song et al. “Denoising Diffusion Implicit Models.” arXiv:2010.02502 (2020)

 

 

(2) Forward Process

2) Jiaming Song et al. “Denoising Diffusion Implicit Models.” arXiv:2010.02502 (2020)

 

 

(3) Trainable Generative Process

2) Jiaming Song et al. “Denoising Diffusion Implicit Models.” arXiv:2010.02502 (2020)

 

 

(4) Loss in Traininable Generative Process

2) Jiaming Song et al. “Denoising Diffusion Implicit Models.” arXiv:2010.02502 (2020)

 

 

(5) Sampling from Generalized Generative Process

2) Jiaming Song et al. “Denoising Diffusion Implicit Models.” arXiv:2010.02502 (2020)

 

 

(6) The Reason Why This Models is 'Implicit'

2) Jiaming Song et al. “Denoising Diffusion Implicit Models.” arXiv:2010.02502 (2020)  / 3) Shakir Mohamed et al. ”Learning in Implicit Generative Models. arXiv:1610.03483(2016)

 

 

(7) DDPM vs. DDIM

2) Jiaming Song et al. “Denoising Diffusion Implicit Models.” arXiv:2010.02502 (2020)

 

 

 

 

출처
1. Jiaming Song et al. “Denoising Diffusion Implicit Models.” arXiv:2010.02502 (2020)
2. Shakir Mohamed et al. ”Learning in Implicit Generative Models.” arXiv:1610.03483(2016)