CNS Sampler boosts diffusion image detail without extra steps
New noise injection method reallocates energy to unfinished frequency bands.
Diffusion models generate images with a spectral bias: low-frequency global structure appears early, while high-frequency detail emerges only at the end. Standard SDE solvers waste stochastic energy by injecting uniform white noise at every step, even into bands that are already resolved. Researchers from the project introduce Colored Noise Sampling (CNS), which reconsiders inference as a targeted energy transfer. At each step, CNS measures how "built" each frequency band is using a precomputed progress index γ(f, t) ∈ [0, 1], and dynamically routes noise energy toward bands with the largest remaining structural deficit. A strict global variance-conservation constraint (mean β² = 1) ensures the modified SDE still converges to the target data distribution.
CNS is a strictly plug-and-play sampler substitution: same model, same number of steps, only the noise injection changes. Early results show sharper textures and finer details compared to standard samplers like DDIM or Euler, especially in complex regions like fur, foliage, or skin pores. The method works with any pre-trained diffusion model (including Stable Diffusion and latent diffusion variants) and requires no additional training or fine-tuning. The project page includes comparisons and the GitHub repository provides an implementation compatible with popular frameworks like diffusers.
- CNS uses a progress index γ(f,t) to measure frequency band completeness and allocate noise energy accordingly.
- The method preserves global variance (mean β² = 1) to maintain convergence guarantees without extra steps.
- Works as a drop-in replacement for any diffusion model sampler, improving high-frequency detail with zero retraining.
Why It Matters
Sharper AI-generated images without extra compute — a simple swap that improves visual quality.