AI Innovation Institute

Abstract: Generative image models trained on massive datasets encode the statistics of our visual world. An off-the-shelf diffusion or flow-matching model can therefore serve as a general-purpose image prior, providing information about images across a wide range of tasks. Harnessing these capabilities, however, requires combining the model at inference time with additional signals or constraints unknown during training. This thesis will develop training-free methods for inference under a generative denoising prior. I first show how sampling from a trained denoiser can be formulated as an optimization problem, which combined with a constraint, can solve tasks ranging from conditional generation and weakly supervised segmentation to combinatorial optimization. I then demonstrate how the inherent properties of denoisers can accelerate inference under such constraints. Replacing gradient descent with an inexact Newton update, based on the symmetry of the denoiser's Jacobian, substantially reduces inference costs without tradeoffs. I also explore a middle ground between model adaptation and fully training-free inference by using the denoiser's robust internal representations to learn constraints from limited labeled data. These methods are applied to gigapixel image domains such as digital histopathology and remote sensing, where generative models can only be trained at a patch scale. To synthesize arbitrarily large images at resolutions unseen during training, I introduce inference-time algorithms that enforce consistency across spatially overlapping patches and image scales. Finally, I propose repurposing inference-time algorithms from sampling tools, to mechanisms for understanding the prior learned by a denoising model. Extending the previously developed techniques, I analyze the Jacobians of generative denoisers, where preliminary results indicate that Jacobian spectra correlates with generative quality. This motivates a Jacobian-spectrum regularization as a way to improve model performance using insights derived from inference-time algorithms.

Speaker: Alexandros Graikos

Location: NCS 220