Search papers, labs, and topics across Lattice.
This paper investigates the geometric underpinnings of phase transitions in continuous-state generative models, specifically focusing on how abrupt qualitative changes in sample trajectories can occur during reverse-time dynamics. By framing denoising as gradient descent on a free energy landscape, the authors identify projection caustics as critical points where unique nearest-point projections fail, leading to sharp transitions. The introduction of the Critical Boundary Detector (CBD) provides a diagnostic tool for identifying mode commitment and intervention-sensitive windows, demonstrating its effectiveness across various generative models.
Abrupt qualitative changes in generative model outputs can be traced to geometric features in the data landscape, revealing critical points that dictate model behavior.
Continuous-state generative samplers, including diffusion and flow-matching models, evolve through continuous reverse-time dynamics, yet their samples often undergo abrupt qualitative changes: trajectories commit to modes, semantic alternatives collapse, and small perturbations in narrow time windows can produce large downstream effects. This paper develops a geometric account of such phase-transition-like behaviour. We view denoising as gradient descent on a free energy landscape and show that sharp transitions arise near projection caustics, where the nearest-point projection onto the data support ceases to be unique. Motivated by this perspective, we introduce the Critical Boundary Detector (CBD), as practical diagnostics for score-direction instability. Across toy models, standard diffusion models, and latent text-to-image diffusion models, CBD localises mode commitment, predicts intervention-sensitive windows, and supports targeted control in geometrically sensitive regions. Our results connect geometry of data and dynamics of diffusion generation.
