Severity-Channel Conditioning: An Injection-Level Study for Consumer-Device Deblurring

Abdullah Al Shafi; Kazi Saeed Alam; Sumaiya Rahim Suma; Sk. Imran Hossain

eess.IVOct 17, 2026

Severity-Channel Conditioning: An Injection-Level Study for Consumer-Device Deblurring

Authors: Abdullah Al Shafi, Kazi Saeed Alam, Sumaiya Rahim Suma, Sk. Imran Hossain

Organizations: Department of Computer Science and Engineering, Khulna University of Engineering & Technology, Khulna 9203, Bangladesh

Abstract

Motion-deblurring models lose substantial accuracy when transferred from professional cameras to smartphone-derived imagery. This paper studies whether a severity descriptor, partly derived from the paired sharp target, helps adaptation on the iPhoneBlur consumer-device benchmark. Under a shared update budget, input concatenation of a severity map is compared with global feature modulation by capture metadata, spatial feature modulation by the severity map, and a tier-contrastive output objective, together with an unconditioned spectral-loss control, all fine-tuned from the same domain-adapted latent-diffusion baseline. The primary experiment is an oracle diagnostic: target-derived information is supplied at inference. For the evaluated model pair, input concatenation improves overall peak signal-to-noise ratio from 31.23 to 31.30 dB, whereas feature modulation, the contrastive objective, and the spectral-loss control do not exceed the baseline. The gain is positive on 87.2% of test images and on 14 of 15 held-out source videos; this video-level analysis describes the evaluated checkpoint pair and does not establish robustness to retraining, and three training seeds are reported separately. The input-level gain recurs on a transformer backbone. In one exploratory run, zeroing the channel at inference retains most of the gain in peak signal-to-noise ratio but lowers structural similarity below the baseline. A normalized-step diagnostic does not support the expected larger-gradient explanation. The results characterize the evaluated configurations and do not establish injection location as the sole cause, a spatial-locality mechanism, or a reliable deployment advantage.

Figures & tables

Fig. 1: Injection levels considered. Dashed arrows show the restorer’s data path, from input image through features and prediction to the loss. Each configuration modifies only its own level, and no experiment combines mechanisms. Input: severity map concatenated to the image. Feature: affine modulation by the metadata vector (global) or the severity map (spatial). Output: a tier-contrastive term on a projection of the prediction. Objective: a fixed-weight spectral loss that receives no descriptor, reported as an unconditioned control.

Fig. 2: SCC integrated with DiffIR. (a) A severity map $\mathbf{s}$ is built from $\mathbf{x}$ and $m_{\text{isp}}$ (Eq. ( 2 )) and concatenated to $\mathbf{x}$ ; the prior encoder’s first projection is widened with zero-initialized weights, $W^{\prime}[:,48{:}64]=\mathbf{0}$ , the only modified layer, so the network equals the DAB at initialization (Proposition 1 ). (b) CPEN S2 encodes $[\mathbf{x};\mathbf{s}]$ into a compact prior, refined by a DDPM reverse process in latent space to give $\hat{\mathbf{z}}$ . (c) The unchanged DIRformer restores $\mathbf{x}$ , conditioned at every block by $\hat{\mathbf{z}}$ , with a global residual connection. The gradient term of $\mathbf{s}$ uses $\mathbf{x}$ alone; $m_{\text{isp}}$ requires paired data (Section III ).

Fig. 3: Severity maps across tiers (rows: Easy, Medium, Hard). Columns: blurred input, ground truth, severity map $\mathbf{s}$ (Eq. ( 2 )), and $\mathbf{s}$ overlaid on the input. Each example is the test image closest to its tier’s median high-frequency difference statistic. The map is high in low-gradient regions and scaled globally by $\gamma(m_{\text{isp}})$ .

Method	GoPro	HIDE	RealBlur-J	RealBlur-R	iPhoneBlur
NAFNet [ 6 ]	$33.7$	$31.3$	$28.5$	$35.1$	$26.6$
HINet [ 23 ]	$32.9$	$30.2$	$27.2$	$34.2$	$26.8$
Restormer [ 24 ]	$32.9$	$31.2$	$29.0$	$35.4$	$26.8$
FFTformer [ 7 ]	$34.0$	$31.6$	$29.2$	$35.6$	$26.9$
DiffIR [ 5 ] ∗	$33.20$	$31.55$	–	–	$27.22$

TABLE I: Zero-Shot Transfer of GoPro-Pretrained Models (PSNR, dB). Baseline values are from [ 1 ] ; DiffIR’s GoPro and HIDE values are from [ 5 ] (RealBlur not reported) and its iPhoneBlur value is measured here. ∗ Evaluated here, not a new model.

Method	Easy	Med.	Hard	Overall	SSIM	LPIPS
MIMO-UNet [ 22 ]	$33.9$	$29.7$	$25.7$	$29.3$	$0.920$	$0.075$
Instruct-IR [ 32 ]	$34.5$	$29.9$	$25.4$	$29.5$	$0.924$	$0.071$
HINet [ 23 ]	$34.6$	$30.9$	$27.1$	$30.5$	$0.938$	$0.056$
Restormer [ 24 ]	$34.9$	$31.2$	$27.5$	$30.9$	$0.940$	$0.053$
FFTformer [ 7 ]	$35.0$	$31.3$	$27.7$	$31.0$	$0.943$	$0.050$
NAFNet [ 6 ]	$35.1$	$31.6$	$27.9$	$31.2$	$0.945$	$0.049$

TABLE II: Fine-Tuned Comparison on iPhoneBlur (PSNR, dB, per Tier; SSIM and LPIPS Overall). Baselines are published benchmark results [ 1 ] , three-seed means, fine-tuned from GoPro weights. The DAB row is seed 1; its three-seed mean is

31.22

dB (Section VI-F ).

Level	Mechanism	Easy	Med.	Hard	Overall	$\Delta$	SSIM	LPIPS
–	Schedule-matched DAB	$35.14$	$31.62$	$27.99$	$31.23$	–	$0.948$	$0.049$
Feat.	FiLM, global (near-identity init)	$34.90$	$31.28$	$27.66$	$30.91$	$-0.32$	$0.945$	$0.053$
Feat.	FiLM, spatial (SFT-style)	$35.12$	$31.59$	$27.98$	$31.20$	$-0.03$	$0.947$	$0.050$
Out.	Contrastive	$35.04$	$31.49$	$27.87$	$31.11$	$-0.12$	$0.946$	$0.051$
Ctrl.	Spectral-loss control (no descriptor)	$35.10$	$31.59$	$28.02$	$31.21$	$-0.02$	$0.944$	$0.052$
In.	SCC (ours)	$\mathbf{35.20}$	$\mathbf{31.70}$	$\mathbf{28.08}$	$\mathbf{31.30}$	$\mathbf{+0.07}$	$\mathbf{0.949}$	$\mathbf{0.048}$

TABLE III: Conditioning Study on iPhoneBlur (PSNR, dB). SSIM and LPIPS are overall;

\Delta

is the overall PSNR change. The descriptor is supplied at inference (oracle). The spectral-loss control receives no descriptor.

Fourth-channel content	Information	PSNR	$\Delta$
None (schedule-matched DAB)	–	$31.23$	–
Random constant	none	$31.24$	$+0.01$
Shuffled severity	global (shuffled)	$31.22$	$-0.01$
Image gradient only	spatial only	$31.22$	$-0.01$
Scale factor $\gamma$ only	global (uniform)	$31.23$	$0.00$
True severity ( SCC )	global + spatial	$\mathbf{31.30}$	$\mathbf{+0.07}$

TABLE IV: Descriptor Ablations (Overall PSNR, dB). All runs share SCC ’s architecture, recipe, and budget and differ only in the fourth channel; single runs at seed 1. The baseline’s three seeds span

[31.18,31.26]

Backbone	Model	Easy	Med.	Hard	Overall	$\Delta$
DiffIR (diffusion)	DAB	$35.14$	$31.62$	$27.99$	$31.23$	–
DiffIR (diffusion)	SCC	$35.20$	$31.70$	$28.08$	$31.30$	$+0.07$
NAFNet (conv.)	Publ.	$35.1$	$31.6$	$27.9$	$31.2$	–
NAFNet (conv.)	SCC	$35.26$	$31.72$	$28.03$	$31.31$	–
Restormer (transf.)	DAB	$34.92$	$31.32$	$27.72$	$30.95$	–
Restormer (transf.)	SCC	$35.00$	$31.42$	$27.80$	$31.04$	$+0.09$

TABLE V: Cross-Backbone Comparison (PSNR, dB; Oracle Descriptor). Each SCC model is compared with its backbone’s baseline. The NAFNet baseline (Publ.) is the published three-seed benchmark result [ 1 ] , shown at its published precision for context only; no matched difference is computed for it.

Tier	$N$	Med. $\Delta$ (dB)	Win %	$p$ -value	95% CI (dB)	$r$
Easy	$280$	$+0.051$	$86.8$	$<10^{-33}$	–	$0.74$
Medium	$953$	$+0.079$	$87.9$	$<10^{-108}$	–	$0.76$
Hard	$453$	$+0.093$	$86.1$	$<10^{-50}$	–	$0.72$
Overall	$\mathbf{1686}$	$\mathbf{+0.079}$	$\mathbf{87.2}$	$\mathbf{<10^{-188}}$	$\mathbf{[0.067,0.073]}$	$\mathbf{0.74}$

TABLE VI: Per-Image Significance of SCC Over the Schedule-Matched DAB on iPhoneBlur (DiffIR backbone).

p

-value is a two-sided Wilcoxon signed-rank test;

r

is the rank-biserial correlation. Oracle descriptor at inference; frame-level, see Section VI-F for the video-clustered check.

Model	Seed	PSNR	SSIM	LPIPS
DAB	1	$31.23$	$0.948$	$0.049$
	2	$31.18$	$0.947$	$0.050$
	3	$31.26$	$0.949$	$0.048$
	mean	$31.22{\scriptstyle\,\pm 0.04}$	$0.948{\scriptstyle\,\pm 0.001}$	$0.049{\scriptstyle\,\pm 0.001}$
SCC	1	$31.30$	$0.949$	$0.048$
	2	$31.34$	$0.952$	$0.050$

TABLE VII: Multi-Seed Comparison on iPhoneBlur (Overall; Oracle Descriptor). Three seeds per arm; only data order and augmentation differ. Paired test over seeds:

p=0.093

(Section VI-F ).

Fig. 4: Mean PSNR gain (

\textsc{SCC}-\text{DAB}

, oracle descriptor) per held-out video, sorted by effect size;

n

is the number of frames per video. The shaded band is the

95\%

equal-video bootstrap CI. Fourteen of

15

videos favour SCC .

DiffIR backbone
Mechanism	Params (M)	Added (M)	FLOPs (G)	Lat. (ms)
Schedule-matched DAB	$26.912$	–	$3380.5$	$3397$
FiLM, global	$28.593$	$1.681$	$3380.5$	$3454$
FiLM, spatial (SFT)	$30.962$	$4.05$	$4926.5$	$4870$
Contrastive	$26.912$	$0$	$3380.5$	$3397$
Spectral-loss control	$26.912$	$0$	$3380.5$	$3397$

TABLE VIII: Inference Cost at

720\times 1280

on One GPU. The contrastive head and the spectral loss are used only in training, so those rows equal the DAB. The lower block covers the other two backbones (Section VI-E ).

Fig. 5: Qualitative comparison on iPhoneBlur. Columns: blurred input, SCC restoration, ground truth; two magnified crops accompany each example (orange and cyan boxes). Rows span the difficulty spectrum.

Field	Formula / Definition	Units	Provenance
Paired high-frequency difference ( $m_{\text{isp}}$ , key isp_diff )	high-frequency energy (mean absolute Laplacian of luminance, as defined by the benchmark [ 1 ] ) of the blurred observation $\mathbf{x}$ minus that of the sharp image $\mathbf{y}$ , min–max normalized over all released metadata, training and test (range $[-3830.9,-33.6]$ ; transductive)	difference of Laplacian energies	y-derived
Image gradient term $g(\mathbf{x})$	Gaussian-smoothed Sobel gradient-magnitude map of $\mathbf{x}$ (Eq. (2) of the main text); high in low-gradient regions, blurred or naturally smooth	dimensionless (gradient magnitude)	x-only
Optical-flow motion magnitude ( motion )	Mean Farnebäck dense optical flow magnitude averaged over the synthesis window: $m_{i}=\frac{1}{W_{i}}\sum_{t=t_{s}}^{t_{e}}\mathbb{E}_{(x,y)}[\lVert v_{t}(x,y)\rVert]$ , where $v_{t}$ is dense flow between consecutive source frames [ 1 , 8 ]	px/frame (range $0.0$ – $28.1$ )	adjacent-frame
Temporal blur window ( blur_window )	Adaptive number of consecutive high-frame-rate source frames averaged to synthesize each blurred sample [ 1 ]	frames (range $3$ – $21$ )	synthesis-only

TABLE IX: Metadata Descriptor Field Provenance

Mechanism	Grad. norm (normalized) †	Grad. SNR	Output change (intensity, $[0,255]$ )
Input ( SCC )	$1.25\times 10^{-5}$	$0.128$	$0.039$
Feature (FiLM)	$2.35\times 10^{-2}$	$0.131$	$21.637$

TABLE X: Normalized-Step Diagnostic: Input Channel vs. FiLM at the Shared DAB Checkpoint

Tier	PSNR	SSIM	LPIPS
Easy	$35.100$	$0.9710$	$0.0310$
Medium	$31.589$	$0.9455$	$0.0500$
Hard	$28.020$	$0.9253$	$0.0703$
Overall	$31.213$	$0.9443$	$0.0523$

TABLE XI: Spectral-Loss Control (No Descriptor), Full Breakdown (PSNR dB / SSIM / LPIPS)

Condition	PSNR (dB)	Max. abs. pixel diff.	SSIM / LPIPS
True severity channel	$31.30$	–	$0.949$ / $0.048$
Zeroed (this audit)	$31.28$	$0$	$0.944$ / $0.048$
Dropped (this audit)	$31.28$	$0$	$0.944$ / $0.048$

TABLE XII: Zero-vs-Drop Determinism Audit

Video	Mean $\Delta$ (dB)	$n$ (frames)
1	$-0.0033$	$55$
2	$+0.0002$	$40$
3	$+0.0168$	$121$
4	$+0.0168$	$72$
5	$+0.0452$	$212$
6	$+0.0573$	$33$

TABLE XIII: Per-Video Mean PSNR Difference and Frame Count, Sorted by Effect Size. The final row is the frame-weighted mean; the equal-video mean is

+0.065

dB.

Parameter	Value
Optimizer	AdamW ( $\beta_{1}=0.9$ , $\beta_{2}=0.99$ , no weight decay)
LR schedule	cosine from $2\times 10^{-5}$ (minimum $10^{-6}$ ), period $34{,}284$ steps, stepped once per optimizer update; final rate $\approx 1.05\times 10^{-5}$
EMA decay	$0.999$
Training stages	DAB adaptation: $12$ epochs from GoPro-pretrained DiffIR; branch stage: $12$ further epochs from the DAB for every DiffIR arm, including the schedule-matched baseline; Restormer: $20$ epochs from GoPro-pretrained weights for both runs
Optimizer steps	$2{,}857$ micro-batches per epoch, accumulation $2$ ; the loss is divided by the accumulation factor under AMP gradient scaling, and the trailing micro-batch of each epoch is discarded, giving $1{,}428$ updates per epoch and $17{,}136$ per DiffIR branch
Batch size	physical $2$ , gradient accumulation $2$ , effective $4$ ; contrastive arm: physical $4$ , no accumulation (same $1{,}428$ updates per epoch, final incomplete batch dropped)

TABLE XIV: Full Training Configuration

Fig. 6: Qualitative comparison with the DAB, omitted from Fig. 5 of the main text for width. Columns: DAB, SCC , ground truth; two magnified high-motion crops per example (cyan, orange). Rows span the difficulty tiers.

Fig. 7: Failure cases, in the layout of Fig. 6 . Blur is concentrated in a small, fast-moving foreground region against a sharp background, and neither model recovers the subject. A single per-image scale $\gamma(m_{\text{isp}})$ cannot represent such localized severity.

Image Deblurring

Image deblurring restores sharp images from motion, defocus, and other blur by estimating lost detail and modeling image formation.

See all 61 papers

iPhoneBlur: A Difficulty-Stratified Benchmark for Consumer Device Motion Deblurring
May 7, 2026Abdullah Al Shafi, Kazi Saeed AlamImage Deblurring Benchmark Design
Department of Computer Science and Engineering Khulna University of Engineering & Technology, Bangladesh
Controllable blind deblurring with diffusion models
Aug 24, 2026Imane Si Salah, Emile Cribelier, Thomas Veit +2Image Deblurring Image Restoration
DxO Labs, France · LTCI, Télécom Paris, IP Paris, France
SloMoDeblur: A Large-Scale Smartphone Image Deblurring Dataset
Jun 24, 2025Syed Mumtahin Mahmud, Mahdi Mohd Hossain Noki, Prothito Shovon Majumder +4Image Deblurring
Cognitive Agents and Interaction Lab, Department of Computer Science and Engineering, University of Dhaka, Dhaka-1000, Bangladesh

May 7, 2026J/K move · Enter open · S save

iPhoneBlur: A Difficulty-Stratified Benchmark for Consumer Device Motion Deblurring

Abdullah Al Shafi, Kazi Saeed Alam

Department of Computer Science and Engineering Khulna University of Engineering & Technology, Bangladesh

Image Deblurring Benchmark Design

Open paper AskPDF arXiv

Motion blur restoration on consumer mobile devices is typically evaluated using aggregate metrics that obscure performance variation across blur difficulty, masking model behavior under real deployment conditions. This work introduces iPhoneBlur, a difficulty-stratified benchmark of 7,400 image pairs synthesized from high-framerate iPhone 17 Pro videos captured in diverse real-world scenarios. Samples are partitioned into Easy, Medium, and Hard categories through PSNR-guided adaptive temporal windowing, with stratification validated by monotonic 2.2x increase in optical flow magnitude across tiers. Each sample includes comprehensive metadata enabling investigation of ISP-aware and difficulty-adaptive restoration strategies. Spectral analysis confirms synthesized blur exhibits high-frequency suppression patterns consistent with authentic motion degradation. Evaluation of six architectures reveals consistent 7-9 dB performance degradation from Easy to Hard subsets, a substantial gap entirely hidden by aggregate reporting. The benchmark further exposes a domain gap between professional and consumer cameras which targeted fine-tuning substantially recovers. By coupling difficulty stratification with deployment-critical metadata, iPhoneBlur enables systematic assessment of model reliability and failure modes for resource-constrained edge systems.

Abstract

Image Deblurring

Explore similar work

iPhoneBlur: A Difficulty-Stratified Benchmark for Consumer Device Motion Deblurring

Controllable blind deblurring with diffusion models

SloMoDeblur: A Large-Scale Smartphone Image Deblurring Dataset