The Geometry of Noise: Why Diffusion Models Don't Need Noise Conditioning
8 hours ago
- #diffusion models
- #generative AI
- #machine learning
- Diffusion models traditionally receive noise level as input to guide denoising.
- Recent research shows noise-level input can be removed for certain models; the geometry of high-dimensional data inherently encodes noise information.
- The posterior distribution over noise levels sharpens in high dimensions, allowing a blind model to match the performance of noise-conditioned models.
- Blind models optimize a single static field, derived from averaging conditioned fields weighted by the posterior probability of noise levels.
- Stability during sampling depends on the model's parameterization: velocity-based models remain stable, while noise-prediction models can diverge when blind.
- The marginal energy landscape provides a static potential for blind models, with singularities canceled by a gain factor to ensure finite updates.