Flow-matching models can produce high-quality image restorations, but their learned dynamics may amplify small input perturbations and make reconstructions unstable. We introduce LipsFlower to reduce this amplification. We derive stability bounds for Flow-from-Y, which initializes the ordinary differential equation from a noise-perturbed observation, as well as for observation-conditioned posterior sampling, which conditions the velocity field on the observation throughout integration. Based on this analysis, we develop LipHDiT, a Lipschitz-aware hourglass diffusion transformer that serves as the vector-field architecture of LipsFlower. LipHDiT incorporates stabilized approximate spectral normalization, bounded modulation, and convex residual mixing. We also use adversarial finite-radius gain regularization at sampled interpolation states to penalize excessive sensitivity to the state and, for conditional sampling, the observation. We test both formulations on blind face restoration, measuring adversarial robustness and the perceptual quality of the reconstructions. LipsFlower reduces perturbation amplification while keeping restoration quality competitive with the evaluated baselines. These results suggest that Lipschitz-aware flow matching can improve the stability of generative image restoration.
Figures & tables
Figure 1: Robustness to observation perturbations. LipsFlower reduces output changes for both backbones under model-specific black-box SPSA attacks, with a larger reduction for Flow-from- Y . The mean perturbation and output differences are amplified for visibility. Blue boxes mark the largest above-threshold change regions, with thresholds matched within each backbone pair.
Figure 2: Flow-based restoration formulations. (a) Flow-from- Y initializes the ODE with the noise-perturbed observation. (b) Observation-conditioned flow starts from a latent z0 independent of y and conditions the velocity field on y throughout integration. Solid arrows between states denote ODE integration, and dashed arrows indicate observation conditioning.
Figure 3: Finite-radius gain regularization. Adversarial searches perturb local state regions and, for conditional sampling (marked in orange), the whole observation. Squared hinge penalties on velocity gains above prescribed thresholds are added to the flow-matching loss.
Method
SPSA– ℓ2 (RMS budget)
SPSA– ℓ∞
Mean G↓
P95 G↓
CVaR95G↓
Max G↓
LPIPS drift ↓
Mean G↓
P95 G↓
CVaR95G↓
Max G↓
LPIPS drift ↓
(I)
CodeFormer
2.44
4.04
4.577
7.17
63.375
2.91
4.82
5.495
9.35
53.101
GFPGAN
3.01
5.13
5.664
7.53
110.573
3.60
6.45
7.122
9.57
98.214
RestoreFormer++
1.77
3.75
4.445
6.79
44.146
2.10
4.56
5.406
8.53
38.113
OSDFace
4.99
6.49
6.898
8.13
186.731
5.83
7.77
8.262
9.87
155.168
PMRF
0.47
1.11
1.455
2.91
9.033
0.47
1.03
1.355
2.96
5.454
Table 1: Adversarial robustness under SPSA attacks on CelebA-Test. LPIPS drift is reported in units of 10−3 . CVaR95G is the mean of the worst 5% among all test images. The blocks (I) and (II) present the comparison sets for the Flow-from- Y and observation-conditioned formulations of LipsFlower, respectively. Within each category, bold and underlined values indicate the best and second-best results, respectively.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Method
PSNR ↑
LPIPS ↓
FID ↓
Precision ↑
DEG ↓
LMD ↓
(I)
CodeFormer
25.146
0.343
52.521
0.3512
37.282
2.470
GFPGAN
24.986
0.364
46.160
0.3212
36.036
2.444
RestoreFormer++
25.313
0.344
40.611
0.4089
29.634
2.043
OSDFace
23.381
0.364
43.567
0.4631
33.907
2.158
PMRF
25.821
0.367
35.842
0.3855
31.027
2.037
ELIR
25.817
0.374
39.944
0.2441
34.258
2.308
Appendix
Table 2: Clean reconstruction and perceptual quality on CelebA-Test at 512×512 resolution. The blocks (I) and (II) present the comparison sets for the Flow-from- Y and observation-conditioned formulations of LipsFlower, respectively. Within each category, bold and underlined values indicate the best and second-best results, respectively.
Figure 4: Qualitative comparison on CelebA-Test. Both LipsFlower variants closely reproduce the facial structure and hairstyle of the ground truth, with differences mainly in fine texture. Despite the trade-off in perceptual quality for improved robustness, their reconstructions remain visually comparable to those of the other methods.
Scale
Probability
Radius distribution
Small
0.70
LogUniform(0.001,0.01)
Moderate
0.20
LogUniform(0.01,0.10)
Severe
0.10
r=0.25
Appendix
Table 3: Mixture used to sample the RMS magnitude inside the perturbed region.
Setting
Flow-from- Y
Conditional
Number of GPUs
4
4
Batch size per GPU
8
16
Gradient accumulation
8
4
Effective batch size
256
256
Regularization interval I
8
8
Selected examples per GPU
2
2
Appendix
Table 4: Default batching and regularization settings.
Figure 5: Additional qualitative comparisons on CelebA-Test. Panels (a)–(c) show three degraded faces and their clean references alongside reconstructions from both LipsFlower variants and the other methods. The examples illustrate how closely LipsFlower recovers facial structure and finer details such as hair and skin texture.
Setting
Backbone
Gain reg.
SPSA– ℓ2 (RMS budget)
SPSA– ℓ∞
Mean G
P95 G
CVaR95G
Max G
LPIPS drift
Mean G
P95 G
CVaR95G
Max G
LPIPS drift
A: Baseline
HDiT
–
0.525
0.743
0.849
1.740
15.803
0.523
0.737
0.830
1.660
9.188
B: Regularizer only
HDiT
Yes
0.559
0.791
0.904
1.584
14.040
0.558
0.770
0.878
1.631
8.660
C: Architecture only
LipHDiT
–
0.486
0.688
0.775
1.309
12.530
0.489
0.674
0.750
1.277
7.772
E: LipsFlower
LipHDiT
Yes
0.475
0.667
0.747
1.158
11.649
0.480
0.658
0.727
1.118
7.360
Appendix
Table 5: Component ablation for Flow-from- Y under SPSA attacks on CelebA-Test. A and B use standard HDiT, without and with sensitivity regularization, respectively; C uses LipHDiT without regularization; E combines both components (full LipsFlower). Lower is better: bold and underlined values indicate the best and second-best point estimates, respectively.
Setting
Backbone
Gain reg.
PSNR ↑
LPIPS ↓
FID ↓
Precision ↑
DEG ↓
LMD ↓
A: Baseline
HDiT
–
25.784
0.350
37.050
0.3797
30.602
1.995
B: Regularizer only
HDiT
Yes
25.298
0.384
35.306
0.4005
31.145
2.045
C: Architecture only
LipHDiT
–
25.474
0.379
39.906
0.2625
34.148
2.345
E: LipsFlower
LipHDiT
Yes
25.475
0.388
40.769
0.2711
34.143
2.351
Appendix
Table 6: Clean reconstruction and perceptual quality for the Flow-from- Y component ablation on CelebA-Test. A and B use standard HDiT, without and with sensitivity regularization, respectively; C uses LipHDiT without regularization; E combines both components (full LipsFlower). Bold and underlined values indicate the best and second-best point estimates, respectively.
Released v1.3 checkpoint with the clean architecture; restoration weight 0.5 ; aligned inputs; upscale factor 1 .
RestoreFormer++ (c)
Released RestoreFormer++ checkpoint; aligned inputs; upscale factor 1 .
OSDFace (d)
Stable Diffusion 2.1-base backbone; one U-Net evaluation at timestep 399 ; merged LoRA with rank 16 and scale parameter 16 ; FP32 inference.
DiffBIR (e)
Version 2 face-restoration pipeline; spaced sampler with 50 steps; classifier-free guidance scale 4 ; captioning disabled; FP32 inference.
PMRF (f)
Official implementation and pretrained checkpoint; 100 flow integration steps.
Appendix
Table 7: Inference configurations for the literature baselines. Iterative methods retain their respective sampling procedures.
Setting
Flow-from- Y
Conditional
Image resolution
5122
5122
Number of GPUs
4
4
Batch size per GPU
8
16
Accumulation steps
8
4
Effective batch size
256
256
Learning rate
5×10−4
5×10−4
Appendix
Table 8: Reference training configuration for the two LipsFlower formulations. Batch size denotes the number of examples per GPU.
Figure 6: Output changes under adversarial perturbations on CelebA-Test. For three degraded faces, panels (a)–(c) show the amplified absolute difference between each method’s reconstructions before and after a model-specific black-box SPSA attack on the observation. Consistent with the quantitative results, LipsFlower reduces output changes for both backbones, with a larger reduction for Flow-from- Y .
Image restoration is fundamentally constrained by the tradeoff between distortion and perception: minimizing pixel-wise error yields over-smoothed results, whereas optimizing for perceptual realism often introduces structural deviations. Recent approaches attempt to balance this tradeoff via posterior sampling or multi-stage generative pipelines, yet remain computationally expensive and architecturally complex. To overcome these limitations, we propose PCFlow (Perceptually Consistent Flow Matching), a unified framework that directly parameterizes a continuous transport from degraded observations to clean targets, jointly optimizing distortion and perceptual quality. While its latent consistency flow objective drives stable and efficient few-step inference, a Latent Consistency Perceptual Loss (LCPL) imposes semantic constraints directly on the guiding velocity field, steering the dynamics toward visually sharp data manifolds. Furthermore, recognizing the inherent conflict between structural and perceptual consistencies, we integrate a conflict-free gradient projection strategy to stabilize the multi-objective optimization landscape. Combined with lightweight, convolution-only backbone, PCFlow achieves competitive performance across diverse restoration tasks at a fraction of traditional computational costs.
Sangwoo Jo, Donggeun Ko, Jayeon Kang +3
Korea University, Seoul, South Korea · Aim Future, Seoul, South Korea
Generative flow models offer powerful priors learned from large-scale natural images, but directly adapting them to restoration tasks such as motion deblurring causes severe fidelity degradation, as their training objective is inherently misaligned with restoration. We present DeblurFlow, a framework that resolves this misalignment by reformulating the flow trajectory itself: we replace the noise endpoint with the blur observation, which makes the underlying vector field coincide with the residual error between blur and clean images. Under this formulation, the standard flow matching loss naturally takes the form of a residual loss, allowing pretrained flow models to be optimized under restoration-aligned objectives via LoRA adaptation. This formulation further enables a dual-expert sampling strategy: a fidelity expert provides a high-fidelity initialization, e.g., PSNR 33.69 dB, and DeblurFlow enhances perceptual quality with only a marginal fidelity reduction to 33.05 dB, whereas directly applying a generative model on top of a fidelity expert decreases PSNR to 27.60 dB. To make this practical, we further introduce r-space, a latent space tailored for residual decoding rather than image reconstruction, which reduces encoder-decoder cost by up to 9×over standard VAE latents. Extensive experiments on GoPro, HIDE, RealBlur, and RWBI demonstrate that DeblurFlow achieves strong restoration fidelity and perceptual realism, while remaining computationally practical.
Current approaches for restoration of degraded images face a trade-off: high-performance models are slow for practical use, while fast models produce poor results. Knowledge distillation transfers teacher knowledge to students, but existing static feature matching methods cannot capture how modern transformer architectures dynamically generate features. We propose a novel Latent Rectified Flow Feature Distillation method for restoring degraded images called 'RestoRect'. We apply rectified flow to reformulate feature distillation as a generative process where students learn to synthesize teacher-quality features through learnable trajectories in latent space. Our framework combines Retinex decomposition with learnable anisotropic diffusion constraints, and trigonometric color space polarization. We introduce a Feature Layer Extraction loss for robust knowledge transfer between different network architectures through cross-normalized transformer feature alignment with percentile-based outlier detection. RestoRect achieves better training stability, and faster convergence and inference while preserving restoration quality, demonstrating superior results across 15 image restoration datasets, covering 4 tasks, on 10 metrics against baselines.