2025 01 Mila

AdjointDEIS: Efficient Gradients for Diffusion Models
Mila – Quebec AI Institute
_{Abstract. Training-free guided generation is a powerful technique to exert additional control on the generative pipeline of diffusion models. We discuss the application of the continuous adjoint equations (sometimes called the adjoint method) on diffusion models. We then use these equations to estimate the gradients of the solution trajectory and conditional information with respect to some differentiable guiding function defined on the output. We discuss special optimizations we can make for Variance Preserving (VP) type diffusions to further simplify the continuous adjoint equations. We demonstrate an application of this technique as an adversarial attack against a Face Recognition (FR) system in the form of a face morphing attack, a powerful biometric attack which attempts to create a single image which successfully matches with multiple bona fide identities.}