Search papers, labs, and topics across Lattice.
This paper investigates the relationship between symmetry breaking and nonlocality phase transitions in diffusion models, two proposed explanations for the critical time window during generation. By analyzing generation trajectories, the authors find that the critical times for non-locality (failure of local denoising) and symmetry breaking (bifurcation into semantic minima) occur nearly simultaneously. This unification provides a practical diagnostic tool for understanding when and why diffusion models require conditioning and global denoising.
Diffusion models' reliance on global information isn't just a quirk – it's fundamentally linked to the moment they commit to a specific semantic outcome.
Diffusion models undergo a phase transition in a critical time window during generation dynamics, with two complementary diagnoses of criticality. The symmetry breaking picture views the critical window as when trajectories bifurcate into different semantic minima of the energy landscape, whereas the nonlocality picture views the critical window as when local denoising fails. We study whether two notions of such phase transitions are concurrent in modern diffusion transformers. By evaluating the dynamics and outcomes of the generation trajectory, we observe a near-simultaneous occurrence of the non-locality and symmetry breaking critical times. Our work is the first to unify the two notions of phase transitions in practice: it provides a concrete diagnostic for when and why diffusion models rely on conditioning and global denoising, enabling principled evaluation of model efficiency and guiding the design of architectures and sampling schemes that avoid unnecessary computation.
