The Noise-Transmission Burden of Diffusion Network: Why Image Prediction Can Outperform Noise Prediction
Abstract
Clean-image prediction (x-prediction) and noise prediction yield equivalent population-optimal denoisers, yet can perform very differently in finite diffusion networks. For noisy input Yt with noise standard deviation σt, a noise predictor fε,t induces the clean-image estimate X̂1=Yt−σt fε,t(Yt). This reveals a noise-transmission burden: the predictor must preserve information needed to cancel the input corruption. Under Gaussian corruption of d-dimensional data, a linear input bottleneck of rank r<d imposes a denoising-error floor σt2(d−r) even with unrestricted downstream computation, while the corresponding optimal risk for image prediction remains bounded. EDM attenuates this bottleneck contribution at high noise; exact preconditioning cancels it at every noise level. Without an input bottleneck, assuming exact affine recovery of the noisy patch from the model representation, we characterize how the covariance spectrum of corrections beyond linear denoising determines the width required above the patch dimension. ImageNet experiments with JiT show that measured denoising losses reflect the predicted bottleneck floor and its attenuation or cancellation through preconditioning. Further widening yields larger relative denoising gains at low noise, where corrections beyond linear denoising spread across more directions, consistent with our spectral characterization of the required model width. Together, these results provide a denoising-based account of practical generation-quality gaps across prediction targets, preconditioning choices, and model widths.
Related
- A forecast-to-dispatch workflow for day-ahead operation of a regional natural gas transmission network based on station-level load prediction and linepack target setting
- Probabilistic AI models outperform numerical weather prediction models on record breaking heat extremes
- Spatial Prediction with Guided Diffusion
- A parallel convolution attention network for crop genomic prediction
- Interpretable Social Network Churn Prediction via Ergodic Markov Chains
BibTeX
@misc{kang2026noisetransmission,
title = {{The Noise-Transmission Burden of Diffusion Network: Why Image Prediction Can Outperform Noise Prediction}},
author = {Gyeonghun Kang and Pengxi Liu and Bingshuo Qian and Xiang Cheng},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.34834},
eprint = {2610.34834},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.34834},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - The Noise-Transmission Burden of Diffusion Network: Why Image Prediction Can Outperform Noise Prediction
AU - Gyeonghun Kang
AU - Pengxi Liu
AU - Bingshuo Qian
AU - Xiang Cheng
PY - 2026
DA - 2026/10/26
PB - arXiv
DO - 10.48550/arxiv.2610.34834
UR - https://arxiv.org/abs/2610.34834
AB - Clean-image prediction (x-prediction) and noise prediction yield equivalent population-optimal denoisers, yet can perform very differently in finite diffusion networks. For noisy input Y_t with noise standard deviation σ_t, a noise predictor f_ε,t induces the clean-image estimate X̂_1=Y_t−σ_t f_ε,t(Y_t). This reveals a noise-transmission burden: the predictor must preserve information needed to cancel the input corruption. Under Gaussian corruption of d-dimensional data, a linear input bottleneck of rank r<d imposes a denoising-error floor σ_t^2(d−r) even with unrestricted downstream computation, while the corresponding optimal risk for image prediction remains bounded. EDM attenuates this bottleneck contribution at high noise; exact preconditioning cancels it at every noise level. Without an input bottleneck, assuming exact affine recovery of the noisy patch from the model representation, we characterize how the covariance spectrum of corrections beyond linear denoising determines the width required above the patch dimension. ImageNet experiments with JiT show that measured denoising losses reflect the predicted bottleneck floor and its attenuation or cancellation through preconditioning. Further widening yields larger relative denoising gains at low noise, where corrections beyond linear denoising spread across more directions, consistent with our spectral characterization of the required model width. Together, these results provide a denoising-based account of practical generation-quality gaps across prediction targets, preconditioning choices, and model widths.
ER -