Tuning Real-World Image Restoration at Inference: A Test-Time Scaling Paradigm for Flow Matching Models

Purui Bai; Junxian Duan; Pin Wang; Jinhua Hao; Ming Sun; Chao Zhou; Huaibo Huang

cs.CVMar 23, 2026

Tuning Real-World Image Restoration at Inference: A Test-Time Scaling Paradigm for Flow Matching Models

Authors: Purui Bai, Junxian Duan, Pin Wang, Jinhua Hao, Ming Sun, Chao Zhou, Huaibo Huang

Organizations: MAIS & NLPR, Institute of Automation, Chinese Academy of Sciences, China · Kuaishou Tech, China

Abstract

Although diffusion-based real-world image restoration (Real-IR) has achieved remarkable progress, efficiently leveraging ultra-large-scale pre-trained text-to-image (T2I) models and fully exploiting their potential remain significant challenges. To address this issue, we propose ResFlow-Tuner, an image restoration framework based on the state-of-the-art flow matching model, FLUX.1-dev, which integrates unified multi-modal fusion (UMMF) with test-time scaling (TTS) to achieve unprecedented restoration performance. Our approach fully leverages the advantages of the Multi-Modal Diffusion Transformer (MM-DiT) architecture by encoding multi-modal conditions into a unified sequence that guides the synthesis of high-quality images. Furthermore, we introduce a training-free test-time scaling paradigm tailored for image restoration. During inference, this technique dynamically steers the denoising direction through feedback from a reward model (RM), thereby achieving significant performance gains with controllable computational overhead. Extensive experiments demonstrate that our method achieves state-of-the-art performance across multiple standard benchmarks. This work not only validates the powerful capabilities of the flow matching model in low-level vision tasks but, more importantly, proposes a novel and efficient inference-time scaling paradigm suitable for large pre-trained models.

Figures & tables

Figure 1: ResFlow-Tuner delivers superior performance on both synthetic (the first row) and real-world (the second row) benchmarks, excelling in terms of perceptual quality and objective image quality assessment.

Figure 2: Architecture of the proposed ResFlow-Tuner . ResFlow-Tuner enhances training performance through the seamless integration of multi-modal guidance. During inference, it adopts a greedy optimization strategy for path selection, augmented by our Multi-Step Partial Denoising Estimator (MSPDE) for more accurate path evaluation.

Panel of Figure 3 — Figure 3: Qualitative comparisons on both synthetic (the first row) and real-world (the last three rows) benchmarks. Please zoom in for a better view.

Dataset	Metric	Real-ESRGAN	BSRGAN	StableSR	DiffBIR	PASD	SeeSR	DreamClear	OSEDiff	SUPIR	FaithDiff	PnP-Flow	FlowSR	LP	Ours
Dataset	Metric	[ 50 ]	[ 61 ]	[ 49 ]	[ 26 ]	[ 57 ]	[ 53 ]	[ 2 ]	[ 52 ]	[ 60 ]	[ 7 ]	[ 30 ]	[ 54 ]	[ 46 ]
DIV2K-Val-I	PSNR $\,{\color[rgb]{1,0,0}\uparrow}$	26.64	27.63	24.71	24.60	25.31	25.08	23.76	25.13	25.09	24.29	23.82	25.02	22.98	23.11
	SSIM $\,{\color[rgb]{1,0,0}\uparrow}$	0.7737	0.7897	0.7131	0.6595	0.6995	0.6967	0.6574	0.7098	0.7010	0.6668	0.6632	0.7219	0.6932	0.6945
	LPIPS $\,{\color[rgb]{1,0,0}\downarrow}$	0.1964	0.2038	0.2393	0.2496	0.2370	0.2263	0.2259	0.2382	0.2139	0.2187	0.2214	0.2143	0.2198	0.2098
	PaQ-2-PiQ $\,{\color[rgb]{1,0,0}\uparrow}$	73.85	74.17	74.19	72.81	73.12	73.92	73.86	74.07	74.24	74.76	73.21	74.14	73.82	78.82
	MANIQA $\,{\color[rgb]{1,0,0}\uparrow}$	0.5611	0.5507	0.5743	0.5898	0.5524	0.5749	0.5750	0.5767	0.6049	0.5945	0.5511	0.5828	0.5842	0.7261

Table 1: Quantitative comparison with state-of-the-art methods on synthetic benchmark. ‘I’, ‘II’, and ‘III’ denote the datasets with mild, medium, and severe degradations, respectively. The best and second performances are marked in red and blue , respectively.

Dataset	Metric	Real-ESRGAN	BSRGAN	StableSR	DiffBIR	PASD	SeeSR	DreamClear	OSEDiff	SUPIR	FaithDiff	PnP-Flow	FlowSR	LP	Ours
Dataset	Metric	[ 50 ]	[ 61 ]	[ 49 ]	[ 26 ]	[ 57 ]	[ 53 ]	[ 2 ]	[ 52 ]	[ 60 ]	[ 7 ]	[ 30 ]	[ 54 ]	[ 46 ]
RealPhoto60	PaQ-2-PiQ $\,{\color[rgb]{1,0,0}\uparrow}$	69.04	63.38	64.40	71.88	69.97	73.77	75.84	73.41	74.42	75.27	72.78	73.82	73.44	78.38
	MANIQA $\,{\color[rgb]{1,0,0}\uparrow}$	0.5094	0.3759	0.4456	0.6253	0.5290	0.6057	0.6166	0.5989	0.6165	0.6527	0.6176	0.6348	0.6320	0.6425
	MUSIQ $\,{\color[rgb]{1,0,0}\uparrow}$	59.29	45.46	57.89	63.67	64.53	70.80	70.46	70.44	70.26	72.74	70.08	71.82	70.96	74.81
	CLIPIQA+ $\,{\color[rgb]{1,0,0}\uparrow}$	0.4389	0.3397	0.4214	0.4935	0.4786	0.5691	0.5273	0.5720	0.5528	0.5932	0.5795	0.5842	0.5856	0.7861
RealSR	MANIQA $\,{\color[rgb]{1,0,0}\uparrow}$	0.3656	0.3660	0.3465	0.4182	0.3855	0.5189	0.4337	0.4052	0.4296	0.4893	0.4910	0.5232	0.5308	0.6205

Table 2: Quantitative comparison with state-of-the-art methods on real-world benchmark. The best and second performances are marked in red and blue , respectively.

Metrics	GT	LQ	Real-ESRGAN [ 50 ]	BSRGAN [ 61 ]	StableSR [ 49 ]	DiffBIR [ 26 ]	SeeSR [ 53 ]	DreamClear [ 2 ]	SUPIR [ 60 ]	FaithDiff [ 7 ]	Ours
Precision	52.72%	7.54%	13.19%	12.04%	19.87%	26.21%	30.07%	22.45%	31.78%	36.45%	42.23%
Recall	56.67%	7.59%	13.68%	12.33%	20.31%	27.90%	33.09%	23.50%	41.57%	46.74%	49.75%

Table 3: Recognition results on the dataset of Occluded RoadText 2024 [ 44 ] (OCR recognition). The best and second performances are highlighted in red and blue , respectively.

	LPIPS ${\color[rgb]{1,0,0}\downarrow}$	DISTS ${\color[rgb]{1,0,0}\downarrow}$	FID ${\color[rgb]{1,0,0}\downarrow}$	MANIQA ${\color[rgb]{1,0,0}\uparrow}$	MUSIQ ${\color[rgb]{1,0,0}\uparrow}$	CLIPIQA ${\color[rgb]{1,0,0}\uparrow}$	Infer-Compute ${\color[rgb]{1,0,0}\downarrow}$	Parameters ${\color[rgb]{1,0,0}\downarrow}$
LR w/o Text Caption	0.3856	0.1468	20.02	0.6542	72.98	0.7212	/	/
LR w/ Text Caption	0.3721	0.1406	19.82	0.6658	74.21	0.7389	/	/
Direct Adding	0.4223	0.1847	28.39	0.5796	69.14	0.6863	/	612M / 5.1%
ControlNet	0.4056	0.1663	24.73	0.6172	71.20	0.6917	/	3.3B / 27.5%
UMMF	0.3721	0.1406	19.82	0.6658	74.21	0.7389	/	58.0M / 0.5%
Aesthetic	0.3872	0.1662	23.25	0.6224	72.57	0.7128	/	/

Table 4: Ablation results on LSDIR-Val for ResFlow-Tuner.

Figure 4: User Study Results. (a) Average ranking of the six methods across all participants and test images, with error bars. (b) Top-K ratios (K=1,2,3,4,5) demonstrating our method’s consistency in producing high-quality results across diverse image content.

Panel of Figure 5 — Figure 5: Visual comparisons for ablation study on ResFlow-Tuner (1/2).

Panel of Figure 6 — Figure 6: Visual comparisons for ablation study on ResFlow-Tuner (2/2).

Panel of Figure 7 — Figure 7: Visual comparisons on real-world benchmarks (1/2). Please zoom in for a better view.

Panel of Figure 8 — Figure 8: Visual comparisons on real-world benchmarks (2/2). Please zoom in for a better view.

Panel of Figure 9 — Figure 9: Visual comparisons on synthetic benchmarks (1/2). Please zoom in for a better view.

Panel of Figure 10 — Figure 10: Visual comparisons on synthetic benchmarks (2/2). Please zoom in for a better view.

Image Restoration

Image restoration uses computational methods to recover useful visual content from noisy, blurred, low-resolution, or otherwise degraded images.

See all 372 papers

ScaleResfusion: Residual Rectified Flow based on Residual Vector Field
Jul 28, 2026Zhenning Shi, Chen Xu, Junhao Zhang +4Diffusion Models for Image Restoration Rectified Flow
College of Computer Science, Nankai University, Tianjin, China · Faculty of Information Science and Computing, University of Macau, Macau, China · CSIRO Data61, Australia +1
In-Token Learning for High-Fidelity Image Restoration via Diffusion Transformers
Sep 27, 2026Xingfu Yi, Xiaoxue YuImage-to-Image Translation Diffusion Transformer
Independent Researcher, Hangzhou, China · Zhejiang University, Hangzhou, China
Multi-Resolution Flow Matching: Training-Free Diffusion Acceleration via Staged Sampling
Jul 2, 2026Xingyu Zheng, Xianglong Liu, Yifu Ding +4Diffusion Model Acceleration Diffusion Model Inference Acceleration
State Key Laboratory of Complex & Critical Software Environment, Beihang University · School of Computer Science and Engineering, Beihang University · Nanyang Technological University +5

Jul 28, 2026J/K move · Enter open · S save

ScaleResfusion: Residual Rectified Flow based on Residual Vector Field

Zhenning Shi, Chen Xu, Junhao Zhang, Kefei Zhang+3

College of Computer Science, Nankai University, Tianjin, China · Faculty of Information Science and Computing, University of Macau, Macau, China · CSIRO Data61, Australia+1

Diffusion Models for Image Restoration Rectified Flow Diffusion Model Fine-Tuning Diffusion-Based Inverse Problems Image Super-Resolution Diffusion-Based Denoising

Open paper AskPDF arXiv

Real-world Image Restoration (Real-IR) aims to recover high-quality (HQ) images from complex and unknown degradations. Recent diffusion-based methods have substantially improved perceptual quality, yet two obstacles remain: methods that sample from Gaussian noise require many steps and are often less faithful to the degraded input, whereas residual-based methods that start from the low-quality (LQ) image typically train task-specific models from scratch, with optimization objectives coupled to a particular noise scheduler, and therefore cannot reuse modern pre-trained generative priors. We present ScaleResfusion, which rewrites residual restoration as a scheduler-independent adaptation interface for pre-trained text-to-image rectified-flow models. Its core, Residual Rectified Flow (RRF), inserts the residual term

R

into the linear transport path of Rectified Flow, so that sampling starts from noisy LQ at an exact acceleration point, where the signal-to-noise ratio of the starting state is continuously controlled by the residual ratio

γ

. The resulting optimization target, the residual vector field, contains no scheduler-specific coefficients and differs from the pre-trained rectified-flow target only by the residual offset

γR

; adapting a frozen billion-scale backbone therefore reduces to fitting this compact residual correction with LoRA-only training. A knowledge-distillation pipeline built around RRF further reduces sampling to as few as 4 steps. Experiments on real-world super-resolution across multiple benchmarks show that ScaleResfusion achieves state-of-the-art restoration quality and transfers consistently across pre-trained rectified-flow backbones from 2B to 9B parameters.

Abstract

Image Restoration

Explore similar work

ScaleResfusion: Residual Rectified Flow based on Residual Vector Field

In-Token Learning for High-Fidelity Image Restoration via Diffusion Transformers

Multi-Resolution Flow Matching: Training-Free Diffusion Acceleration via Staged Sampling