The Geometry of Noise: Why Diffusion Models Don't Need Noise Conditioning

A team of researchers from Google (Mojtaba Sahraee-Ardakan, Mauricio Delbracio, and Peyman Milanfar) has published a breakthrough theoretical paper titled 'The Geometry of Noise: Why Diffusion Models Don't Need Noise Conditioning' on arXiv. The work resolves a fundamental paradox in autonomous generative models like Equilibrium Matching and blind diffusion, which learn a single, time-invariant vector field without explicit noise-level conditioning. The central question was how these bounded, noise-agnostic networks remain stable near the data manifold where gradients typically diverge.

The researchers formalized the concept of 'Marginal Energy' (E_marg(u) = -log p(u)), where p(u) is the marginal density of noisy data integrated over unknown noise levels. They proved that generation using autonomous models isn't simple blind denoising, but rather a specific form of Riemannian gradient flow on this Marginal Energy landscape. Through a novel relative energy decomposition, they demonstrated that while the raw Marginal Energy has a 1/t^p singularity normal to the data manifold, the learned time-invariant field implicitly incorporates a local conformal metric that perfectly counteracts this geometric singularity, converting an infinitely deep potential well into a stable attractor.

Crucially, the paper identifies why certain parameterizations fail while others succeed. They found a 'Jensen Gap' in noise-prediction parameterizations that acts as a high-gain amplifier for estimation errors, explaining catastrophic failures in deterministic blind models. Conversely, velocity-based parameterizations are inherently stable because they satisfy a bounded-gain condition that absorbs posterior uncertainty into smooth geometric drift. This theoretical framework provides the mathematical foundation for understanding and designing more stable, efficient diffusion models that don't require explicit noise conditioning.