@teropa: I which @sedielem beautifully illustrates why diffusion models work so well with images Our visual world is spatially c…
Summary
An explanation of why diffusion models work well for images: low-frequency spectral components dominate, so denoising recovers coarse structure first, then fine detail — analogous to spectral autoregression.
View Cached Full Text
Cached at: 06/17/26, 01:43 AM
I which @sedielem beautifully illustrates why diffusion models work so well with images
Our visual world is spatially coherent, and large scale structures dominate. 〜 This means low frequency spectral components in images tend to be stronger. This means when you noise up an image, high frequency detail goes away first, then coarse structure. This means when you denoise an image, the coarse structure emerges first, then the fine detail. This means you can think of diffusion as “spectral autoregression”. Coarse / low end to fine / high end.
Similar Articles
Show the Signal, Hide the Noise: Spectral Forcing for Pixel-Space Diffusion
A new technique called Spectral Forcing applies a time-conditional 2D-DCT low-pass operator to pixel-space diffusion models, improving efficiency by explicitly separating signal from noise and outperforming baselines on ImageNet and text-to-image tasks.
@docmilanfar: I really enjoyed the explainer for our recent paper on "Geometry of Noise" arXiv:2602.18428
This paper provides a theoretical explanation for why diffusion models can generate clean samples without explicit noise-level conditioning, attributing it to high-dimensional geometry and analyzing why some model parameterizations succeed while others collapse.
Deep Dreams Are Made of This: Visualizing Monosemantic Features in Diffusion Models
This paper introduces Latent Visualization by Optimization (LVO), a mechanistic interpretability technique that uses sparse autoencoders to visualize monosemantic features in diffusion models like Stable Diffusion 1.5.
Decoupled Residual Denoising Diffusion Models for Unified and Data Efficient Image-to-Image Translation
This paper proposes Decoupled Residual Denoising Diffusion Models (DRDD) for unified and data-efficient image-to-image translation, decoupling noise diffusion for domain harmonization from residual diffusion for semantic mapping.
Spectral Guidance for Flexible and Efficient Control of Diffusion Models
Introduces Spectral Guidance, a framework for controlling diffusion models by leveraging low-dimensional representations of the diffusion process, enabling flexible and stable control without task-specific retraining or backpropagation through the denoiser.