Dynamic Time Step Prediction in Inverse Heat Dissipation for Blur-Like Image Restoration Tasks

Cap Dang Xuan Kiet; Tat-Jen Cham

cs.CVOct 2, 2026

Dynamic Time Step Prediction in Inverse Heat Dissipation for Blur-Like Image Restoration Tasks

Authors: Cap Dang Xuan Kiet, Tat-Jen Cham

Organizations: Nanyang Technological University Singapore

Abstract

When using diffusion models to target image restoration problems, diffusion inversion is typically employed to retain relevant image information from the degraded images. Instead of inverting back to the initial time step (i.e., T), many methods invert to a pre-determined intermediate time step, in order to better preserve information from degraded source images. However, a pre-determined time step for inversion is not ideal for reconstruction, as a severely degraded image requires an earlier starting time step than a mildly degraded one. In addition, DDIM-based models corrupt the original signal by adding Gaussian noise, which can be mismatched to the nature of blur-like degradations, such as blur, haze, and low-light. To address these problems, we propose two solutions: (1) we adopt an alternative diffusion process, called the Inverse Heat Dissipation Model, that diffuses the input image by gradually blurring a data point (2) we propose to implement a time predictor to estimate the starting time step for the inversion, with the model learning to adapt to the degradation severity. Extensive experiments on standard benchmarks show that our method achieves state-of-the-art performance in both quantitative and qualitative evaluations, with excellent generalization to many restoration tasks.

Figures & tables

Figure 1: Real-world applications of our method. We use images from the realistic dataset RealBlur [ 42 ] for the deblurring task and internet images for the dehazing and low-light enhancement tasks. Our method can generate high-fidelity results.

Figure 2: Image restoration by diffusion inversion methods. $(a)$ In inversion-based methods, the input image $y_{0}$ is first mapped to the noise space by the forward process. Then the extracted noise is used to solve the reverse process. $(b)$ [ 62 ] perturbs the model with additive noise, aligning $y_{0}$ to an intermediate $x_{s}$ . This method is feasible with a vast number of reverse problems whose input is only slightly deviated from the expected output. $(c)$ Our method aims to find the optimal time step $s$ , determined by the degree of degradation, at which $y_{s}$ and $x_{s}$ are highly similar. A reverse network is trained to revert $y_{s}\approx x_{s}$ to $x_{0}$ , adapting to multiple severities.

Figure 3: Inference flow of our proposed method.

Models	Low-Light					Haze					Blur					#Param
Models	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	KID $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	KID $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	KID $\downarrow$	M
InvSR	18.48	0.809	0.159	35.91	10.56	29.36	0.930	0.041	39.27	5.95	28.02	0.799	0.049	57.36	18.87	33.8
DFPIR	18.74	0.843	0.146	35.12	10.92	29.63	0.932	0.043	30.17	3.22	28.17	0.813	0.055	56.26	14.44	31.1
MoCE-IR	20.04	0.889	0.124	34.43	9.32	30.16	0.978	0.031	29.01	1.92	28.64	0.857	0.051	56.58	15.37	25.4
FoundIR	21.40	0.989	0.004	34.37	8.92	30.85	0.964	0.034	28.98	2.26	28.37	0.869	0.045	54.24	13.18	36.3
Our model	19.98	0.883	0.127	33.92	9.13	30.70	0.983	0.021	28.84	2.02	29.09	0.880	0.033	53.99	13.01	37.7

Table 1: Quantitative comparisons to SOTAs in the all-in-one setting. KID values are scaled by a factor of 100 for readability. Red , Blue , and Orange denote the first, second, and third best, respectively.

Models	Low-light
Models	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FID $\downarrow$	KID $\downarrow$
Diff LL	20.47	0.879	0.143	37.36	13.31
PNP	13.83	0.649	0.492	60.06	28.49
InstructP2P	14.33	0.757	0.398	34.31	17.90
Diff-Plugin	19.95	0.888	0.121	33.26	7.28
InvSR	17.85	0.853	0.145	34.37	9.09

Table 2: Quantitative comparison in single-task setting. From left to right: Low-light enhancement, dehazing, and debluring.

Methods	Ours	Diff-Plugin	InstructP2P	PNP
Avg Inference Time (s)	1.626	2.994	8.891	0.724

Table 3: Average inference time comparison on a single A100 GPU with the input size of

512\times 512

Figure 4: Qualitative comparison. From left to right: degraded image, ground truth, ours, Diff-Plugin [ 31 ] , PNP [ 50 ] , DFPIR [ 48 ] , and expert models. Please zoom in for better visualization.

Model selection	MI $\uparrow$
Model selection	RealBlur	LOLI-Street
DM [ 46 ]	0.915	0.257
BDM [ 16 ]	0.921	0.325
IHDM [ 43 ]	1.021	0.327

Table 4: Mutual Information (MI) between

x_{t}

and

y_{t}

at the optimal time steps. We test our proposed method using different diffusion models as the backbone and the same optimization objective.

Model selection	RealBlur		LOLI-Street
Model selection	FID $\downarrow$	KID $\downarrow$	FID $\downarrow$	KID $\downarrow$
DM [ 46 ]	54.98	16.35	33.46	9.77
BDM [ 16 ]	55.10	16.18	33.39	9.42
IHDM [ 43 ]	51.35	11.44	32.90	8.76

Table 5: Comparison between using different models as backbone.

Figure 5: Qualitative results of IHDM on GOPRO dataset [ 35 ] at different time steps.

Figure 6: Quantitative results of IHDM on GOPRO dataset [ 35 ] at different time steps.

Methods	Gaussian Blur (PSNR/SSIM)
Methods	$r=0.5$	$r=1$	$r=2.5$	$r=5$	$r=10$
InvSR	29.88/0.889	25.97/0.791	22.76/0.617	20.37/0.498	18.61/0.401
DFPIR	30.02/0.909	26.36/0.818	23.23/0.663	21.11/0.534	19.18/0.464
Our model	33.24 / 0.948	32.04 / 0.908	29.97 / 0.888	26.78 / 0.814	25.55 / 0.799

Table 6: Quantitative results on the image deblurring task with different levels of blur. In this experiment, in order to easily modify the level of degradation, we apply a synthetic Gaussian blur, which is relatively controlled by the blur radius

r

, on 1221 test images from all categories of ImageNet.

Figure 7: Visual results on GOPRO dataset [ 35 ] . From left to right: input, (a) pre-trained $f_{\theta}$ with pre-defined time steps, (b) train $g_{\phi}$ only, (c) train both $f_{\theta}$ and $g_{\phi}$ , and ground truth.

Figure 8: Mutual Information between $x_{t}$ and $y_{t}$ at different time steps along the diffusion process.

Tasks	w/ TS Predictor	w/o TS Predictor
Tasks	SSIM/LPIPS	SSIM/LPIPS
Blur	0.880/0.033	0.813/0.047
Haze	0.983/0.021	0.909/0.063
Low-light	0.883/0.127	0.789/0.178

Table 7: Ablation study of the effectiveness of the time step predictor.

	$\epsilon=0$	$\epsilon=1$	$\epsilon=3$
	Acc $\uparrow$ / F1 $\uparrow$	Acc $\uparrow$ / F1 $\uparrow$	Acc $\uparrow$ / F1 $\uparrow$
In Dist	96.53/0.9823	98.06/0.9902	98.88/0.9944
Out Dist	79.39/0.8001	82.62/0.8311	85.34/0.8776

Table 8: Results of our time step predictor on predicting the optimal time step.

Diffusion Models for Image Restoration

Diffusion models for image restoration use generative priors to denoise, deblur, upscale, and reconstruct images from incomplete or corrupted measurements.

See all 181 papers

Residual Diffusion Implicit Models
Sep 26, 2026João Guerreiro, Pedro Tomás, Helena Aidos +1Diffusion Models for Image Restoration Diffusion-Based Inverse Problems
INESC-ID, Instituto Superior Técnico, Universidade de Lisboa, Lisboa, Portugal · ISR-Lisboa, Instituto Superior Técnico, Universidade de Lisboa, Lisboa, Portugal · LASIGE, Faculdade de Ciências, Universidade de Lisboa, Lisboa, Portugal
Stop Denoising Your Blurs
May 24, 2026Sasidhar Parvathireddy, Vamsidhar Saraswathula, Rama Krishna GorthiDiffusion Models for Image Restoration Image Deblurring
Indian Institute of Technology Tirupati, India.
Learning Normalized Energy Models for Linear Inverse Problems
May 15, 2026Nicolas Zilberstein, Santiago Segarra, Eero Simoncelli +1Bayesian Inverse Problems Energy-Based Models
Rice University, Houston, TX, USA · Flatiron Institute, New York, NY, USA · New York University, New York, NY, USA

Sep 26, 2026J/K move · Enter open · S save

Residual Diffusion Implicit Models

João Guerreiro, Pedro Tomás, Helena Aidos, Jacinto C. Nascimento

INESC-ID, Instituto Superior Técnico, Universidade de Lisboa, Lisboa, Portugal · ISR-Lisboa, Instituto Superior Técnico, Universidade de Lisboa, Lisboa, Portugal · LASIGE, Faculdade de Ciências, Universidade de Lisboa, Lisboa, Portugal

Diffusion Models for Image Restoration Diffusion-Based Inverse Problems Diffusion Models Image Denoising Diffusion Model Inference Acceleration Image Super-Resolution

Open paper AskPDF arXiv

Diffusion models achieve state-of-the-art results across multiple tasks. However, in inverse problems, standard initialization from pure Gaussian noise misaligns the generative process with real-world degradations. More recent methods such as diffusion bridges impose strict endpoint constraints and often require long reverse processes that are prone to hallucinations. Alternative consistency models provide noise-invariant, one-step mappings but lack inherent variance modeling and can degrade under severe corruption. Hence, residual diffusion implicit models (RDIMs) are proposed, constituting a generalized framework that explicitly models the residuals between high-quality (HQ) and low-quality (LQ) images, aligning the forward process with the actual degradation. A non-Markovian implicit reverse sampler is derived, which can skip intermediate timesteps, enabling accurate few-step or even single-step reconstruction, while mitigating the hallucinations inherent to long diffusion chains. RDIM also introduces a controllable variance mechanism that interpolates between deterministic and stochastic sampling, balancing fidelity and diversity. Furthermore, it enables the straightforward use of perceptual losses, when needed. Experiments on denoising and super-resolution benchmarks demonstrate that RDIMs consistently outperforms the state of the art, including bridge and consistency models, in terms of PSNR, SSIM, and LPIPS, reducing hallucinations while requiring only a few sampling steps (often just one). The results position RDIMs as an efficient solution for a broad range of image restoration tasks.

Abstract

Diffusion Models for Image Restoration

Explore similar work

Residual Diffusion Implicit Models

Stop Denoising Your Blurs

Learning Normalized Energy Models for Linear Inverse Problems