1 of 17

Introduction to Diffusion-based Generative Models

YANG Can

macyang@ust.hk

Department of Mathematics, HKUST

Fall, 2024

2 of 17

Review of Deep Generative models (GAN)

  • Unstable.
  • Mode collapse
  • Density ratio is undefined where no training data is available

 

3 of 17

Review of Deep Generative models (VAE and flow-based)

4 of 17

Score-based methods

  •  

https://yang-song.net/blog/2021/score/

5 of 17

Connection with Tweedie’ formula

  •  

Tweedie’s formula

Herbert Robbins (1956) credits personal correspondence with Maurice Kenneth Tweedie for an extraordinary Bayesian estimation formula.

6 of 17

Naive score-based generative modeling and its pitfalls

1. The manifold hypothesis

The manifold hypothesis states that data in the real world tend to concentrate on low dimensional manifolds embedded in a high dimensional space (a.k.a., the ambient space)

2. Inaccurate score estimation with score matching

In regions of low data density, score matching may not have enough evidence to estimate score functions accurately, due to the lack of data samples.

7 of 17

Naive score-based generative modeling and its pitfalls

8 of 17

Naive score-based generative modeling and its pitfalls

3. Trouble in mode recovery

When two modes of the data distribution are separated by low density regions, Langevin dynamics will not be able to correctly recover the relative weights of these two modes in reasonable time, and therefore might not converge to the true distribution.

[Ref] Luo, C. (2022). Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970.

9 of 17

Score-based generative modeling with multi-scale noise perturbations

  • Multi-scale noise perturbations resolve all the above three challenges.

10 of 17

Score-based generative modeling with multi-scale noise perturbations

 

11 of 17

https://yang-song.net/blog/2021/score/

12 of 17

Cifar10 and Celebra

https://yang-song.net/blog/2021/score/

13 of 17

Diffusion models

 

 

  1. Langevin dynamics

  • Ancestral sampling

 

 

 

 

 

[Ref] Ho, J., Jain, A., & Abbeel, P. (2020). Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33, 6840-6851.

[Ref] Vincent, P. (2011). A connection between score matching and denoising autoencoders. Neural computation, 23(7), 1661-1674.

14 of 17

Denoising diffusion probabilistic models (DDPM)

 

  • DDPM

 

 

 

True Posterior mean (used for solving inverse problems)

 

 

15 of 17

Problem Setting

Linear inverse problem

 

 

Inpainting

Super Resolution

Non-uniform deblur (Non-linear)

[ref] Chung, H., Kim, J., Mccann, M. T., Klasky, M. L., & Ye, J. C. (2022). Diffusion Posterior Sampling for General Noisy Inverse Problems. arXiv preprint arXiv:2209.14687.

16 of 17

https://onlinelibrary.wiley.com/doi/10.1111/insr.12002

17 of 17

Reference

[1] Chung, H., Kim, J., Mccann, M. T., Klasky, M. L., & Ye, J. C. (2022). Diffusion Posterior Sampling for General Noisy Inverse Problems. arXiv preprint arXiv:2209.14687.

[2] Luo, C. (2022). Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970.

[3] Song, Y., & Ermon, S. (2019). Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32.

[4] Ho, J., Jain, A., & Abbeel, P. (2020). Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33, 6840-6851.

[5] Song, J., Meng, C., & Ermon, S. (2020). Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502.

[6] Bao, F., Li, C., Zhu, J., & Zhang, B. (2022). Analytic-dpm: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. arXiv preprint arXiv:2201.06503.

[7] Chung, H., Sim, B., Ryu, D., & Ye, J. C. (2022). Improving Diffusion Models for Inverse Problems using Manifold Constraints. arXiv preprint arXiv:2206.00941.

[8] Song, Y., Shen, L., Xing, L., & Ermon, S. (2021). Solving inverse problems in medical imaging with score-based generative models. arXiv preprint arXiv:2111.08005.

[9] Chung, H., Kim, J., Mccann, M. T., Klasky, M. L., & Ye, J. C. (2022). Diffusion Posterior Sampling for General Noisy Inverse Problems. arXiv preprint arXiv:2209.14687.

[10] Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., & Poole, B. (2020). Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456.

[11] Choi, J., Kim, S., Jeong, Y., Gwon, Y., & Yoon, S. (2021). Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938.

[12] Jalal, A., Arvinte, M., Daras, G., Price, E., Dimakis, A. G., & Tamir, J. (2021). Robust compressed sensing mri with deep generative priors. Advances in Neural Information Processing Systems, 34, 14938-14954.

[13] Wei, X., Fu, S., Li, H., Liu, Y., Wang, S., Feng, W., ... & Gu, Y. (2022). Single-cell Stereo-seq reveals induced progenitor cells involved in axolotl brain regeneration. Science, 377(6610), eabp9444.

[14] Vincent, P. (2011). A connection between score matching and denoising autoencoders. Neural computation, 23(7), 1661-1674.