Noise Frequencies Let a Pixel-Space Diffusion Prior
Detect Native-Resolution Image

Bumsoo Kim^1,*, Hyun-Jun Jin^2,*, Mi Young Lee¹, Seungmin Rho¹, Sanghyun Seo^1,†

¹Chung-Ang University · ²Korea University

^*Now with Smilegate AI · ^†Corresponding author

ECCV 2026 Workshop · Low-Level Vision Frontier (LoViF)

🤗 Demo on Hugging Face SOON Dataset on HF SOON arXiv SOON

Teaser

An image at a given resolution does not always carry the quality that resolution implies. Left: a 32×32 pixel-art can be losslessly recovered from its downscaled version — its true (effective) resolution is 32, well below the stored 256. Right: an artistic portrait shows visible reconstruction error at every downscale level, indicating a native effective resolution near 256. PixRes makes this distinction operational as a training-free native-image detector using a pretrained pixel-space diffusion prior.

TL;DR

PixRes asks whether an image sits at its native resolution or has been resampled from a lower one — a training-free binary check built on a pretrained pixel-space diffusion prior. A single PF-ODE (DDIM) inversion up to T/4 extracts a variance signal s_var over the noise's high-frequency spectrum, giving a pooled detection AUC of 0.953 across 8 diverse domains — no supervision, no fine-tuning.

Input
x₀ 256×256

→

PF-ODE inversion
DDIM, T/4

→

Radial spectrum
R(x₀; t, k)

→

Score s_var
std over T_p×K_HF

→

Detection
native? / r_eff

Headline Results

Pooled AUC

0.953

+0.035 vs SS-ERE

Subtle regime

0.807

+0.079 vs SS-ERE

Acc@10%

65.9%

+8.8 pp vs SS-ERE

Spearman ρ

0.709

+0.194 vs SS-ERE

Cross-domain means over 8 benchmarks — FFHQ, Set5, Set14, BSD100, Urban100, DIV2K, COCO, WikiArt. No labelled resolution data used at any point.

Native vs. Resampled — a Clean Split

(a) Pooled over 7 diverse domains, native images (blue) sit near log₁₀ s_var ≈ 3.2, while resampled ones (red) sit near 3.7 — pooled AUC = 0.960. (b) As the downscaling anchor r_d shrinks from 256 (native) to 64 (heavy downscaling), the distribution shifts monotonically — s_var is a true continuous signal of effective resolution.

Explore the Signal Interactively

Predicted diffusion noise ε_θ(x_t) at timestep t = 150 for each sample. Compare native vs. bicubically resampled inputs. (Timestep slider temporarily locked at t = 150 while we finish the multi-timestep results.)

Sample

Baboon Monarch Lenna Zebra Pepper

Timestep t

t=25t=150t=250

t=150

Predicted noise magnitude | native vs upscaled 128 (bili / bicu) vs upscaled 64 (bili / bicu)

Radial power spectrum R(x₀; t, k) (all t averaged)

Watch the high-frequency region (k ≥ 64) in the spectrum panel: the native curve stays flat, while the resampled curves lift and bulge — exactly the mismatch s_var quantifies.

Method Overview

Deterministic PF-ODE inversion in pixel space. A clean input x₀ is inverted along a deterministic trajectory up to t = T/4. At probe timesteps T_p = {τ₁, …, τ₄} we collect the predicted noise 𝜀̄_θ(τ), take its 2-D FFT, and radially bin to obtain R(x₀; t, k). For a native input (blue path, x_HQ) the high-frequency shell (cyan circle) stays uniformly dim — the prior's expected noise complexity matches the input. For a resampled input (red path, x_LQ) the same shell brightens as the prior tries to compensate for missing HF content. The joint std of R over the plateau and HF band is our scalar score s_var.

The intuition is clean: at native resolution, the diffusion prior's expected noise complexity matches the input's true high-frequency content, so the radial spectrum stays flat across the plateau. When the input has been resampled, the prior still expects rich high-frequency noise, creating a variance mismatch that s_var captures directly.

Test Datasets

Random samples from the eight test benchmarks used throughout our experiments — a deliberate mix of natural, structured, face, and stylised domains. All images are prepared identically (center-crop + Lanczos to 256×256).

Dataset	n	Domain	Role
DIV2K val	100	Natural, HR	calibration anchor source
DIV2K train	100	Natural, HR	held-out evaluation
COCO	100	Natural, in-the-wild	evaluation
BSD100	100	Natural	evaluation
Urban100	100	Structured (buildings)	evaluation
FFHQ	100	Face	SS-ERE’s training domain
WikiArt	100	Stylised (painting)	domain-gap stress test
Set5 / Set14	5 / 14	Classical SR	completeness

Cross-Domain Robustness

Three orthogonal aspects of score quality on seven benchmarks: (a) pooled detection AUC, (b) tight-tolerance r_eff regression accuracy (Acc@10%), (c) Spearman rank correlation. Ours (blue) dominates all three — +0.039 AUC, +11.1 pp Acc@10%, and +0.235 ρ on average — including on FFHQ, the domain on which SS-ERE was trained.

Per-Domain ROC Curves

One panel per benchmark. In the practically relevant low-FPR regime (FPR ≤ 0.10) — the region where a real detection system operates — Ours (blue solid) reaches TPR ≥ 0.9 on COCO / DIV2K / Urban100, while SS-ERE (red dashed) plateaus below 0.85. The two ties are Set14 (n=14) and Set5 (n=5), both small-sample edge cases.

Native-Image Detection AUC

For every domain we ask whether the score can distinguish a native 256×256 image from a bicubically resampled one at r_d ∈ {64, 96, 128, 160, 192, 224}. Pooled AUC averages across the six negative anchors weighted by number of pairs.

Dataset	Method	vs 64	vs 96	vs 128	vs 160	vs 192	vs 224	Pooled
FFHQ	SS-ERE	1.000	1.000	0.999	0.980	0.920	0.814	0.952
FFHQ	Ours	1.000	1.000	1.000	1.000	0.996	0.989	0.998
Set5 (n=5)	SS-ERE	1.000	1.000	1.000	0.960	0.840	0.740	0.923
Set5 (n=5)	Ours	1.000	1.000	1.000	1.000	0.880	0.680	0.927
Set14 (n=14)	SS-ERE	1.000	1.000	1.000	0.990	0.872	0.742	0.934
Set14 (n=14)	Ours	0.995	1.000	0.832	0.985	0.954	0.837	0.934
BSD100	SS-ERE	1.000	0.992	0.982	0.937	0.840	0.719	0.912
BSD100	Ours	0.987	0.998	0.990	0.996	0.952	0.807	0.955
Urban100	SS-ERE	1.000	0.996	0.978	0.899	0.753	0.637	0.877
Urban100	Ours	0.996	0.990	0.983	0.989	0.964	0.695	0.936
DIV2K	SS-ERE	1.000	0.992	0.985	0.953	0.870	0.745	0.924
DIV2K	Ours	1.000	1.000	0.992	0.996	0.962	0.771	0.954
COCO	SS-ERE	1.000	0.983	0.978	0.945	0.857	0.677	0.907
COCO	Ours	0.997	1.000	0.999	1.000	0.995	0.858	0.975
WikiArt	SS-ERE	0.991	0.989	0.975	0.939	0.860	0.750	0.917
WikiArt	Ours	0.971	0.984	0.965	0.985	0.959	0.816	0.947
Mean	SS-ERE	0.999	0.994	0.987	0.950	0.851	0.728	0.918
Mean	Ours	0.993	0.996	0.970	0.994	0.958	0.807	0.953

Blue-highlighted values mark the winner per column per dataset. Both methods trivially handle heavy downscaling (r_d ≤ 128); the real gap opens in the subtle-resampling regime (vs 224: 0.807 Ours vs 0.728 SS-ERE, a +7.9 pp advantage at just 12.5% downscaling).

Extension: Continuous r_eff Prediction

The same s_var score also serves as a regression signal. A piecewise-linear calibration on 7 DIV2K-val anchors turns it into a continuous effective-resolution predictor. Reported on the half-native band r_d ∈ {128, 160, 192, 224, 256}.

Dataset	Method	MAE ↓	MAPE ↓	Acc@10% ↑	Acc@20% ↑	Spearman ρ ↑
FFHQ	SS-ERE	22.6	13.6%	48%	74%	0.84
FFHQ	Ours	25.1	15.5%	51%	72%	0.86
Set5	SS-ERE	28.4	17.5%	40%	60%	0.79
Set5	Ours	27.5	16.7%	40%	64%	0.80
Set14	SS-ERE	31.0	19.3%	36%	57%	0.83
Set14	Ours	29.8	18.8%	43%	67%	0.67
BSD100	SS-ERE	26.8	16.0%	38%	65%	0.76
BSD100	Ours	33.3	19.8%	41%	62%	0.83
Urban100	SS-ERE	35.3	21.5%	30%	55%	0.71
Urban100	Ours	29.2	17.9%	44%	68%	0.80
DIV2K*	SS-ERE	28.0	17.0%	37%	64%	0.79
DIV2K*	Ours	22.8	14.1%	55%	76%	0.84
COCO	SS-ERE	31.8	19.4%	32%	54%	0.79
COCO	Ours	28.6	17.6%	47%	67%	0.87
WikiArt	SS-ERE	24.2	13.8%	44%	73%	0.76
WikiArt	Ours	32.8	20.0%	44%	62%	0.82
Mean	SS-ERE	28.5	17.2%	38%	63%	0.78
Mean	Ours	28.6	17.5%	46%	67%	0.81

Blue-highlighted values = per-metric winner. Ours "clean-sweeps" 3 of 8 domains (Urban100, DIV2K, COCO) and wins Acc@10% + Spearman ρ in the cross-domain mean. SS-ERE remains competitive on absolute error metrics (MAE/MAPE), reflecting its supervised regression objective.

Ablations

(a) Timestep T_p: single-timestep native-vs-resampled AUC rises and saturates by t = T/4 (250), justifying the plateau window and the T/4 inversion truncation. (b) Score statistic: among single-statistic candidates our joint (t,k)-variance s_var beats HF energy and log–log spectral slope. (c) Frequency band: AUC peaks near k_min = H/4 = 64 and collapses beyond k > 100, justifying K_HF = [64, 128).

BibTeX

@inproceedings{kim2026pixres,
  title     = {Noise Frequencies Let a Pixel-Space Diffusion Prior Detect Native-Resolution Image},
  author    = {Kim, Bumsoo and Jin, Hyun-Jun and Lee, Mi Young and Rho, Seungmin and Seo, Sanghyun},
  booktitle = {ECCV 2026 Workshop on Low-Level Vision Frontier (LoViF)},
  year      = {2026}
}

References

SS-ERE. Kansy et al. — Self-supervised effective-resolution estimation with adversarial augmentations. WACV 2023. [link]
ADM. Dhariwal & Nichol — Diffusion models beat GANs on image synthesis. NeurIPS 2021. [link]
DDIM inversion. Song et al. — Denoising diffusion implicit models. ICLR 2021. [link]
Score-based generative modelling / PF-ODE. Song et al. — Score-based generative modeling through stochastic differential equations. ICLR 2021. [link]
DiffusionAD / Inv-AD (inspiration). Mousakhan et al., DAGM 2024; Sakai & Hasegawa, arXiv 2504.
Pixel-space diffusion — frequency-aware structure. Ma et al. (DeCo, 2025); Yu et al. (PixelDiT, 2025).

Noise Frequencies Let a Pixel-Space Diffusion PriorDetect Native-Resolution ImageNoise Frequencies Let a Pixel-Space Diffusion PriorDetect Native-Resolution Image