Diffusion Generative Models in Infinite Dimensions

18 concepts15 papers1 proposal

Standard generative models, from GANs to Diffusion, typically operate on finite-dimensional vectors. Whether it is a $256 \times 256$ pixel grid or a fixed-length audio clip, the model treats the data as a point in $\mathbb{R}^n$. This "discretize-first" approach anchors the model to a specific resolution; changing the sampling rate or the grid size often requires retraining or leads to significant artifacts.

A new paradigm shifts this focus by treating data as continuous functions—elements of infinite-dimensional Hilbert or Banach spaces. By defining the generative process directly in these function spaces, researchers have developed models that are inherently resolution-independent and discretization-invariant.

From Pixels to Functions

In traditional diffusion, we noise an image by adding Gaussian noise to its pixels. In infinite dimensions, we instead noise a function $f$ by evolving it according to a Stochastic Differential Equation (SDE) in a Hilbert space. As Kerrigan et al. describe, this allows the model to represent continuous signals—like weather patterns or audio—independent of how they were originally sampled.

The core intuition is to replace the standard multivariate Gaussian prior with a Gaussian Process (GP) prior. As Mathieu et al. argue, defining the forward noising process toward a GP ensures that the marginal distributions remain consistent across any finite set of spatial coordinates. Whether you sample 10 points or 10,000, they all behave as realizations of the same underlying continuous noisy signal.

The Score in Infinite Dimensions

In finite dimensions, the "score" is the gradient of the log-density, $\nabla \log p(x)$. However, in infinite-dimensional spaces, standard probability densities do not exist. To overcome this, Lim et al. utilize the concept of Fomin differentiable measures, defining a score operator that acts on the Hilbert space.

This mathematical lifting allows the generative process to be formulated as the time-reversal of a stochastic evolution equation. Baldassari et al. show that this reverse drift can be learned using Neural Operators, such as Fourier Neural Operators (FNOs). Unlike standard CNNs, which learn mappings between fixed-size tensors, neural operators learn mappings between function spaces. This is the engine behind discretization invariance: the model learns the physics of the signal rather than the statistics of the grid.

Discretization Invariance and Efficiency

The primary advantage of this approach is the ability to "train once, sample anywhere." Rahman et al. demonstrate with Generative Adversarial Neural Operators (GANO) that models can produce high-resolution outputs from low-resolution training data without fine-tuning. Similarly, Corallo et al. show that Functional Diffusion Processes (FDPs) can achieve high-quality results with significantly fewer parameters than traditional models by focusing on the underlying signal representation.

This flexibility extends to irregular data. Bond-Taylor et al. introduce $\infty$-Diff, which uses mollified (smoothed) states and random coordinate subsampling to scale beyond the resolution of the training data while reducing memory costs. By treating the image as a continuous signal, the model avoids the high-frequency artifacts that typically plague standard architectures when pushed beyond their training distribution.

Solving Inverse Problems

Infinite-dimensional models are particularly powerful for Bayesian inference and physical inverse problems, such as seismic imaging or fluid dynamics. Yao et al. demonstrate that extending Tweedie’s formula to Banach spaces allows for "plug-and-play" guidance, where a single pre-trained functional model can solve various reconstruction tasks without retraining.

To handle complex conditioning, the authors of several papers move beyond simple time-reversal. Pieper-Sethmacher et al. and Baker et al. leverage Doob’s h-transform. Instead of trying to reverse a noising process, they "force" a reference diffusion toward a target distribution. This ensures that the generative process remains well-defined and mathematically rigorous even as the resolution of the data increases toward infinity.

Beyond Diffusion: Flows and Kernels

The shift to infinite dimensions is not limited to diffusion. Zhang et al. extend rectified flows to Hilbert spaces, providing a unified theory for functional generative flows that handle irregularly sampled data. For training Neural SDEs, Issa et al. propose using signature kernel scores, replacing the unstable minimax games of GANs with a principled scoring rule based on path signatures.

By moving the mathematical foundation from vectors to functions, these works provide a robust framework for modeling the continuous nature of the physical world, ensuring that our generative models are as fluid as the phenomena they seek to replicate.

Key Foundational Work

Score-based Generative Modeling through Stochastic Evolution Equations in Hilbert Spaces (2023) — contributed 2 concept elements to this direction.
Geometric Neural Diffusion Processes (2023) — contributed 2 concept elements to this direction.
Continuous-Time Functional Diffusion Processes (2023) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_infinite_dimensional_generative_modeling_778877ff,
  title = {infinite-dimensional generative modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/infinite-dimensional-generative-modeling/rdr_778877ffeebf44adaea429ccf1fbdeb8},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_infinite_dimensional_generative_modeling_778877ff_v1,
  title = {infinite-dimensional generative modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/infinite-dimensional-generative-modeling/rdr_778877ffeebf44adaea429ccf1fbdeb8/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake