Towards a Mechanistic Explanation of Diffusion Model Generalization

Matthew Niedoba, Berend Zwartsenberg, Kevin Patrick Murphy, Frank Wood
Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:46389-46411, 2025.

Abstract

We propose a simple, training-free mechanism which explains the generalization behaviour of diffusion models. By comparing pre-trained diffusion models to their theoretically optimal empirical counterparts, we identify a shared local inductive bias across a variety of network architectures. From this observation, we hypothesize that network denoisers generalize through localized denoising operations, as these operations approximate the training objective well over much of the training distribution. To validate our hypothesis, we introduce novel denoising algorithms which aggregate local empirical denoisers to replicate network behaviour. Comparing these algorithms to network denoisers across forward and reverse diffusion processes, our approach exhibits consistent visual similarity to neural network outputs, with lower mean squared error than previously proposed methods.

Cite this Paper

BibTeX
@InProceedings{pmlr-v267-niedoba25a, title = {Towards a Mechanistic Explanation of Diffusion Model Generalization}, author = {Niedoba, Matthew and Zwartsenberg, Berend and Murphy, Kevin Patrick and Wood, Frank}, booktitle = {Proceedings of the 42nd International Conference on Machine Learning}, pages = {46389--46411}, year = {2025}, editor = {Singh, Aarti and Fazel, Maryam and Hsu, Daniel and Lacoste-Julien, Simon and Berkenkamp, Felix and Maharaj, Tegan and Wagstaff, Kiri and Zhu, Jerry}, volume = {267}, series = {Proceedings of Machine Learning Research}, month = {13--19 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v267/main/assets/niedoba25a/niedoba25a.pdf}, url = {https://proceedings.mlr.press/v267/niedoba25a.html}, abstract = {We propose a simple, training-free mechanism which explains the generalization behaviour of diffusion models. By comparing pre-trained diffusion models to their theoretically optimal empirical counterparts, we identify a shared local inductive bias across a variety of network architectures. From this observation, we hypothesize that network denoisers generalize through localized denoising operations, as these operations approximate the training objective well over much of the training distribution. To validate our hypothesis, we introduce novel denoising algorithms which aggregate local empirical denoisers to replicate network behaviour. Comparing these algorithms to network denoisers across forward and reverse diffusion processes, our approach exhibits consistent visual similarity to neural network outputs, with lower mean squared error than previously proposed methods.} }
Endnote
%0 Conference Paper %T Towards a Mechanistic Explanation of Diffusion Model Generalization %A Matthew Niedoba %A Berend Zwartsenberg %A Kevin Patrick Murphy %A Frank Wood %B Proceedings of the 42nd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2025 %E Aarti Singh %E Maryam Fazel %E Daniel Hsu %E Simon Lacoste-Julien %E Felix Berkenkamp %E Tegan Maharaj %E Kiri Wagstaff %E Jerry Zhu %F pmlr-v267-niedoba25a %I PMLR %P 46389--46411 %U https://proceedings.mlr.press/v267/niedoba25a.html %V 267 %X We propose a simple, training-free mechanism which explains the generalization behaviour of diffusion models. By comparing pre-trained diffusion models to their theoretically optimal empirical counterparts, we identify a shared local inductive bias across a variety of network architectures. From this observation, we hypothesize that network denoisers generalize through localized denoising operations, as these operations approximate the training objective well over much of the training distribution. To validate our hypothesis, we introduce novel denoising algorithms which aggregate local empirical denoisers to replicate network behaviour. Comparing these algorithms to network denoisers across forward and reverse diffusion processes, our approach exhibits consistent visual similarity to neural network outputs, with lower mean squared error than previously proposed methods.
APA
Niedoba, M., Zwartsenberg, B., Murphy, K.P. & Wood, F.. (2025). Towards a Mechanistic Explanation of Diffusion Model Generalization. Proceedings of the 42nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 267:46389-46411 Available from https://proceedings.mlr.press/v267/niedoba25a.html.

Related Material