It’s Never Too Late: Noise Optimization for Collapse Recovery in Trained Diffusion Models

• Mode collapse
• Diffusion models usually initialize from white Gaussian noise, i.e., flat power spectrum
• But natural images obey a \(\frac{1}{f}\) spectrum, and most structure lives in low frequencies
• Authors tackle mode collapse by optimizing the initial noise \(x_0\) at inference time
• Sample a bunch of noise latents, generate images, and then compute diversity metrics like DINOv2, LPIPS, DPP, Vendi