WorksComputer Science
Retrieval-Augmented Initialization with Dual-Guided Inversion for Flow-Based Inverse Problem Solvers
Abstract
Flow models have recently emerged as powerful generative priors for solving inverse problems. Despite their success, most existing flow-based inverse problem solvers initialize the reconstruction process with random Gaussian noise. However, recent studies in image generation have shown that the choice of initial noise can significantly influence generation outcomes. The role of initialization as a distinct design choice in flow-based inverse problem solving remains comparatively underexplored. In this work, we systematically evaluate noise initialization strategies while holding the downstream flow-based inverse solver fixed. We propose Retrieval-Augmented Initialization with Dual-Guided Inversion (RA-DGI), a method that constructs improved initial noise by leveraging retrieved proxy clean images and dual-guided inversion. Extensive experiments across multiple inverse problems demonstrate that the quality of the initial noise substantially affects reconstruction fidelity and perceptual performance. RA-DGI outperforms standard random initialization and achieves superior results compared to inference-time scaling strategies, while requiring less computational overhead. These findings highlight initialization as a critical yet previously overlooked factor in flow-based inverse problem solvers.
Introduction
Diffusion models have become a standard approach for image synthesis [24, 26, 27]. Beyond unconditional image generation, diffusion models have also been successfully applied to solving inverse problems, where they serve as powerful image generative priors [5, 17, 33, 37]. Formally, given a measurement generated from an unknown clean signal through a forward degradation process, the goal is to reconstruct the clean signal that is consistent with the measurement . Recently, flow-based generative models have emerged as an efficient alternative to diffusion models for image generation [9, 19, 20, 34]. As a result, there has been growing interest in developing inverse problem solvers that leverage flow models [8, 18, 25]. Despite these advances, most inverse problem solvers still initialize the reconstruction process with randomly sampled Gaussian noise. This design choice contrasts with recent findings in image generation, where carefully chosen initial noise has been shown to significantly improve generation fidelity [10, 11, 13, 36, 39]. Although initialization has received increasing attention in image generation, its role as a distinct design choice in inverse problem solving remains comparatively underexplored.
To address this limitation, we propose a new method, Retrieval-Augmented Initialization with Dual-Guided Inversion (RA-DGI), to improve the initial noise for flow-based inverse problem solvers. First, given a degraded measurement, we retrieve a proxy clean sample from the dataset that is similar in the measurement space. This retrieved proxy clean sample serves as a more suitable starting point for inversion. Second, during inversion, we consider two complementary principles: (1) the encoded latent should remain close to the Gaussian prior distribution and (2) measurement consistency must be continuously enforced throughout the inversion process. By jointly satisfying these constraints, we obtain an initialization that is both distributionally aligned and measurement-aware.
We demonstrate that the choice of initialization significantly affects reconstruction fidelity. As shown in Figure 1(a), our method produces substantially better reconstructions than random Gaussian initialization. Prior work has shown that inference-time scaling can significantly improve the performance of diffusion models [21]. Motivated by this observation, we investigate whether a similar strategy is effective for inverse problems. Figure 1(b) plots reconstruction performance (PSNR, higher is better) against inference-time computational cost. Even when additional computation is allocated to random initialization, the resulting performance remains inferior to our method. These results suggest that improving initialization is more effective and computationally efficient than simply increasing the amount of inference-time compute.
Figure 1. (a) RA-DGI produces reconstructions that are closer to the ground truth than those obtained with a random Gaussian initialization. (b) Reconstruction performance (PSNR) versus inference-time computational cost (NFE). Even with increased compute, random initialization remains inferior to RA-DGI.
Our contributions are as follows:
We propose RA-DGI to construct measurement-aware initial noise.
We demonstrate that the proposed initialization outperforms random initialization and existing baselines in terms of reconstruction fidelity and perceptual quality.
We show that our method is more computationally efficient than inference-time scaling strategies, achieving superior performance with lower computational cost.
Related Work
Solving Inverse Problems with Diffusion/Flow Models
Inverse problems aim to recover a clean signal from degraded measurements , typically modeled as , where is a known measurement operator and denotes measurement noise. Recent approaches address this ill-posed problem using diffusion priors learned from large-scale image datasets [5, 17, 33, 37].
Diffusion-based inverse solvers guide the reverse process using the measurement-conditioned score , which can be decomposed by Bayes’ theorem as
The first term represents the learned generative prior, while the second enforces measurement consistency. Existing solvers mainly differ in how they approximate and incorporate this measurement-likelihood term [6].
Existing diffusion-based inverse solvers enforce measurement consistency either through likelihood gradients derived from the forward model [5, 17, 33, 37] or through variational approximations of the posterior [2, 12, 22]. Flow-based approaches extend this principle by incorporating measurement guidance into probability-flow dynamics [8, 18, 25]. However, these flow-based solvers typically initialize reconstruction with random Gaussian noise. RA-DGI instead focuses on constructing a more informative initial latent and is complementary to existing flow-based solvers.
Noise Initialization and Source Optimization
Recent studies have shown that initial noise substantially affects the quality and prompt alignment of diffusion-based text-to-image generation [10, 11, 13, 36, 39]. INITNO [13] optimizes initial noise based on attention responses during early diffusion steps, while ReNO [10] uses reward-based optimization to improve perceptual quality. Other approaches amortize this optimization through learned networks: NPNet [39] transforms random noise into prompt-aligned noise, and HyperNoise [11] generates improved noise distributions using a prompt-conditioned hypernetwork. However, these methods primarily target prompt-aligned image generation rather than measurement-conditioned reconstruction.
Initialization has also been studied through source-space optimization for inverse problems with pretrained flow models. D-Flow [3] optimizes the initial source point for each measurement by differentiating through the full generation trajectory. MS-Flow [7] instead optimizes intermediate latent states and stitches local trajectory segments, reducing the memory and numerical-stability issues of full-trajectory differentiation. Nevertheless, both methods require iterative, instance-specific optimization. RA-DGI instead constructs a measurement-aware initial latent through retrieval and dual-guided inversion, avoiding iterative optimization of source variables for each measurement.
Method
Background
Flow Models Flow models [19, 20] aim to transport a simple source distribution (typically an isotropic Gaussian) to a target data distribution . This transport is defined through a time-dependent velocity field that governs the following ordinary differential equation (ODE):
By integrating this ODE from to , a sample is deterministically transported to a data sample . Therefore, generation amounts to estimating the marginal velocity field . To define the velocity field, flow models first construct a probability path that connects and . A common choice is the linear interpolation path:
where and , and typically and . However, the true marginal velocity field corresponding to this path is generally intractable. To address this issue, flow models are trained using the conditional flow matching (CFM) objective [19]:
where denotes the conditional velocity field associated with the chosen probability path. Although the training target is the conditional velocity field, minimizing this objective yields a model that approximates the true marginal velocity field [19]. As a result, the learned vector field defines a probability flow that transports to during inference. In practice, new samples are generated by numerically solving the ODE 1 from to . Using a first-order Euler discretization, the update rule is given by
which progressively transforms an initial Gaussian sample into a data sample through step-wise integration.
Inversion Since the ODE defines a deterministic mapping between the data and latent distributions, it can be solved in both time directions. In particular, given a data sample , we can obtain its corresponding latent representation by integrating the ODE forward from to . Using a first-order Euler discretization, the update rule is given by
which progressively transports the data sample toward its inverted latent code in the Gaussian space. Conversely, starting from and integrating the ODE backward from to recovers the original data sample under sufficiently accurate numerical integration [28].
Tweedie’s Formula for Flow Models Tweedie’s formula provides a way to estimate clean variables from intermediate states through conditional expectations. FlowDPS [18] extends this idea to flow models and shows that both the clean data sample and the latent noise sample can be estimated from an intermediate state using the velocity field. Specifically, the conditional expectations of and given are expressed as
where denotes the marginal velocity field at time . Using these expressions, the velocity field can be decomposed as
Retrieval-Augmented Initialization
A straightforward strategy for obtaining an initial latent code is to directly invert the given measurement . However, flow models are trained on clean images rather than degraded measurements. As a result, directly inverting measurement often produces a latent representation that deviates from the Gaussian prior, which can lead to poor reconstruction quality [28]. To mitigate this issue, we instead retrieve a proxy clean sample from the dataset and use it as the starting point for inversion. Intuitively, a clean image that is consistent with the measurement can provide a more suitable initialization for inversion.
Following the standard formulation of inverse problems, the measurement model is written as
where denotes the forward degradation matrix and represents additive Gaussian noise with variance . Let denote the pseudo-inverse of , which maps measurements back to the image space. For each clean sample , we simulate the same degradation process and map the result back to the image space:
We then extract feature representations of each degraded-clean sample and the measurement using a pretrained encoder :
The proxy sample is selected by minimizing the feature distance:
The retrieved proxy clean image serves as the starting point for inversion. Inverting from yields a latent code that more closely aligns with the Gaussian prior, thereby improving reconstruction performance. Figure 2a illustrates the overall process of retrieval-augmented initialization.

Figure 2. Overview of RA-DGI. (a) Retrieval-augmented initialization. A proxy clean image is retrieved from a dataset using semantic features extracted by and used as the starting point for inversion. (b) Dual-guided inversion. Measurement guidance (blue) enforces measurement consistency, while prior guidance (red) applies Gaussian regularization to encourage compatibility with the source prior, jointly refining the latent trajectory.
Dual-Guided Inversion
Measurement Guidance Because the inversion trajectory is initialized from the proxy clean image, the resulting latent code tends to reconstruct the proxy image rather than satisfy the measurement constraint . Our objective, however, is to obtain a latent representation that is consistent with the measurement. To this end, we perform inversion using a measurement-conditional velocity field instead of the unconditional velocity field . Accordingly, Eq. (5) is modified as follows:
The measurement-conditional velocity field can be expressed as
where and [18]. Substituting this expression into Eq. (13) gives
Applying Tweedie’s formula (Eq. 6) further expands this expression as
Furthermore, by applying the velocity decomposition in Eq. (8), we obtain
where
, , ,
and
By continuously incorporating this measurement-consistency term into the inversion dynamics, the latent trajectory is guided toward regions that are consistent with the measurement.
Prior guidance Focusing solely on measurement guidance may cause the encoded latent to deviate from the Gaussian prior. To address this issue, we further modify the measurement-guided inversion step. The term represents the predicted noise inferred from the latent state , which implicitly contains structural information derived from the measurement. Consequently, directly using in the inversion dynamics may bias the latent toward measurement-specific structures rather than maintaining consistency with the Gaussian prior.
To mitigate this effect, we introduce a noise regularization mechanism that injects Gaussian noise to weaken the structural information contained in . Specifically, we replace with the following noise-refined representation:
where denotes a standard Gaussian sample that is independently resampled at every inversion step. This stochastic mixing acts as a regularizer that reduces persistent measurement-specific structure and encourages compatibility with the Gaussian source prior. With this modification, the final inversion update in Eq. (19) becomes
Figure 2(b) illustrates the overall process of dual-guided inversion.
Experiments
Experimental Settings
Inverse Problems We set the number of function evaluations (NFE) to 10 for all experiments. We evaluate our method on four linear inverse problems: super-resolution, deblurring, denoising, and compressed sensing. For super-resolution, we consider a downscaling setting, while for deblurring, we apply motion blur with a kernel size of 81. For denoising, additive Gaussian noise with a standard deviation of is used. For compressed sensing, we adopt a block-wise compressed sensing scheme with a block size of and a measurement ratio of 0.1. For all inverse problems except denoising, measurement noise with a standard deviation of is added.
Datasets and Metrics We conduct experiments on three datasets: AFHQ [4], FFHQ [16], and DIV2K [1]. From AFHQ, we use 1,000 images from the validation set, and from FFHQ, we select 1,000 images. From DIV2K, we use 800 high-resolution images from the validation set. For quantitative evaluation, we report PSNR and SSIM to measure pixel-level reconstruction accuracy. To assess perceptual quality, we use LPIPS [38] as a perceptual similarity metric and FID [14] to evaluate distributional alignment with the ground-truth data distribution.
Comparison Methods Our method aims to identify an effective noise initialization strategy for flow-based inverse problem solvers. As a basic baseline, we consider the standard random Gaussian initialization, where the latent noise is sampled directly from the prior distribution. Because initialization strategies for inverse problem solving have been comparatively less studied, we construct several relevant baselines for comparison. First, we include approaches originally proposed for identifying initial noise in image generation, particularly NPNet [39]. Second, we consider inversion-based strategies that obtain the initial noise by inverting the measurement. In particular, we evaluate RF-Inversion [28], an inversion method designed for flow-based models. For a fair comparison, all initialization strategies are integrated into the same inverse problem solver, FlowDPS [18], so that performance differences can be attributed to the initialization quality rather than the solver itself.
Implementation Details For retrieval-augmented initialization, we use DINOv3 [29] as the feature encoder to extract semantic features for nearest-neighbor search. For the retrieval dataset, we randomly select 1,000 images from Unsplash [31]. All images are resized to . During inference, the classifier-free guidance [15] scale is set to 2.0. For AFHQ, we use the text prompt “a photo of a close-up face of a cat (or dog)”. For FFHQ, we use the prompt “a photo of a close-up face”. For DIV2K, we generate the text prompt from the degraded image using the degradation-aware prompt extractor (DAPE) introduced in SeeSR [35].
Main Results
Quantitative Results Table 1 summarizes the quantitative comparison across four inverse problems on AFHQ, FFHQ, and DIV2K. For this main comparison, we repeat each experiment using three independent random seeds and report the mean performance, with the standard deviation shown in parentheses. Overall, RA-DGI achieves strong pixel-level reconstruction fidelity across datasets and tasks, attaining the highest mean PSNR in 11 of the 12 settings and the highest mean SSIM in all 12 settings. Improvements in LPIPS and FID vary across settings. Random initialization samples the latent code directly from the Gaussian prior without considering the measurement, which often yields latents inconsistent with the given measurement and thus degrades reconstruction quality. NPNet attempts to identify a more suitable initial noise by aligning the latent with the text prompt. However, because it does not explicitly model the measurement, it often fails to achieve better reconstruction performance than random initialization. RF-Inversion estimates the initial latent by inverting the measurement through the flow model. This strategy partially improves the alignment between the latent code and the target image. However, the inversion process starts from the degraded measurement and does not explicitly enforce measurement consistency during inversion, thereby limiting its performance relative to our approach. In contrast, the proposed RA-DGI, which initializes the inversion from a retrieved proxy clean image and enforces measurement consistency during inversion, results in a latent that is more suitable for solving inverse problems than those produced by the baseline methods.
| Dataset | Method | Super-Resolution: PSNR | Super-Resolution: SSIM | Super-Resolution: LPIPS | Super-Resolution: FID | Deblurring: PSNR | Deblurring: SSIM | Deblurring: LPIPS | Deblurring: FID | Denoising: PSNR | Denoising: SSIM | Denoising: LPIPS | Denoising: FID | Compressed Sensing: PSNR | Compressed Sensing: SSIM | Compressed Sensing: LPIPS | Compressed Sensing: FID |
| AFHQ | Random | 23.34 (0.050) | 0.477 (0.004) | 0.235 (0.004) | 21.91 (0.763) | 22.79 (0.021) | 0.464 (0.003) | 0.254 (0.004) | 24.76 (1.272) | 21.92 (0.015) | 0.446 (0.004) | 0.301 (0.004) | 31.22 (0.720) | 20.25 (0.006) | 0.423 (0.003) | 0.353 (0.004) | 41.72 (1.104) |
| AFHQ | NPNet | 23.18 (0.067) | 0.474 (0.003) | 0.240 (0.007) | 24.41 (1.159) | 22.59 (0.061) | 0.453 (0.003) | 0.261 (0.006) | 26.62 (1.306) | 21.93 (0.060) | 0.446 (0.003) | 0.293 (0.005) | 33.24 (1.330) | 20.35 (0.061) | 0.416 (0.003) | 0.338 (0.003) | 42.17 (1.227) |
| AFHQ | RF-Inversion | <u>23.94</u> (0.064) | <u>0.529</u> (0.004) | 0.287 (0.000) | 22.95 (1.104) | <u>23.22</u> (0.057) | <u>0.486</u> (0.003) | 0.299 (0.004) | 23.83 (0.693) | <u>22.05</u> (0.075) | 0.406 (0.005) | 0.243 (0.003) | 28.27 (0.473) | <u>22.32</u> (0.053) | 0.441 (0.002) | 0.267 (0.007) | 29.97 (0.868) |
| AFHQ | RA-DGI (Ours) | 24.03 (0.101) | 0.537 (0.003) | 0.236 (0.007) | 21.92 (0.768) | 23.57 (0.115) | 0.528 (0.003) | 0.266 (0.006) | 24.69 (0.668) | 22.85 (0.085) | 0.475 (0.002) | 0.266 (0.004) | 28.38 (0.286) | 22.69 (0.025) | 0.454 (0.000) | 0.262 (0.001) | 29.72 (0.229) |
| FFHQ | Random | 23.04 (0.006) | 0.461 (0.005) | 0.210 (0.004) | 62.98 (0.693) | 22.61 (0.085) | 0.447 (0.003) | <u>0.237</u> (0.003) | <u>61.18</u> (0.542) | <u>21.83</u> (0.064) | <u>0.433</u> (0.001) | 0.248 (0.004) | 68.92 (0.730) | 20.45 (0.078) | 0.404 (0.002) | 0.298 (0.007) | 81.98 (0.779) |
| FFHQ | NPNet | 22.28 (0.096) | 0.419 (0.003) | 0.246 (0.003) | 79.00 (1.227) | 21.96 (0.042) | 0.402 (0.002) | 0.271 (0.004) | 79.19 (0.267) | 21.66 (0.069) | 0.393 (0.001) | 0.247 (0.003) | 78.74 (0.934) | 20.67 (0.066) | 0.370 (0.003) | 0.283 (0.004) | 90.22 (1.320) |
| FFHQ | RF-Inversion | <u>23.82</u> (0.057) | <u>0.525</u> (0.005) | 0.246 (0.005) | <u>58.98</u> (0.473) | <u>22.95</u> (0.061) | <u>0.470</u> (0.001) | 0.276 (0.004) | 64.48 (0.730) | 21.43 (0.115) | 0.391 (0.004) | 0.197 (0.003) | <u>62.17</u> (0.705) | 21.64 (0.031) | 0.417 (0.003) | 0.230 (0.004) | 67.61 (1.330) |
| FFHQ | RA-DGI (Ours) | 24.41 (0.064) | 0.570 (0.003) | 0.210 (0.003) | 53.46 (0.509) | 23.88 (0.025) | 0.547 (0.000) | 0.237 (0.001) | 57.47 (0.462) | 22.90 (0.053) | 0.472 (0.003) | 0.218 (0.004) | 64.39 (0.743) | 22.43 (0.060) | 0.433 (0.003) | 0.224 (0.004) | 81.53 (0.798) |
| DIV2K | Random | 18.88 (0.060) | 0.334 (0.005) | 0.370 (0.003) | 96.61 (0.868) | 18.78 (0.062) | 0.344 (0.001) | <u>0.400</u> (0.004) | <u>104.52</u> (0.779) | 19.10 (0.046) | <u>0.382</u> (0.005) | 0.379 (0.000) | 96.57 (1.330) | 17.48 (0.044) | 0.325 (0.003) | 0.464 (0.004) | 115.29 (0.163) |
| DIV2K | NPNet | 18.81 (0.061) | 0.339 (0.003) | 0.410 (0.003) | 109.16 (0.106) | 18.90 (0.064) | 0.342 (0.002) | 0.414 (0.004) | 109.26 (1.316) | 19.14 (0.092) | 0.374 (0.003) | 0.391 (0.004) | 99.44 (0.516) | 17.33 (0.070) | 0.324 (0.003) | 0.467 (0.001) | 115.14 (0.867) |
| DIV2K | RF-Inversion | 19.59 (0.075) | <u>0.381</u> (0.004) | 0.414 (0.004) | <u>87.99</u> (0.823) | 19.42 (0.115) | <u>0.357</u> (0.002) | 0.453 (0.003) | 112.42 (1.050) | 19.16 (0.015) | 0.349 (0.002) | 0.320 (0.003) | 88.03 (0.404) | <u>18.67</u> (0.105) | <u>0.326</u> (0.003) | 0.394 (0.004) | <u>109.16</u> (1.267) |
| DIV2K | RA-DGI (Ours) | <u>19.42</u> (0.105) | 0.392 (0.002) | <u>0.376</u> (0.002) | 86.93 (0.229) | 19.56 (0.080) | 0.398 (0.003) | 0.394 (0.007) | 93.34 (0.697) | 19.81 (0.067) | 0.405 (0.003) | <u>0.346</u> (0.003) | <u>89.67</u> (0.794) | 18.72 (0.076) | 0.349 (0.002) | <u>0.398</u> (0.003) | 107.66 (1.159) |
Table 1. Quantitative comparison on four inverse problems across three datasets. Values are reported as the mean over three seeds, with the sample standard deviation shown in parentheses. Best and second-best mean values are shown in bold and <u>underline</u>, respectively.
Qualitative Results Figure 3 presents qualitative comparisons across four inverse problems. Overall, RA-DGI produces reconstructions that are visually closer to the ground truth than those obtained from other initialization strategies. Random and NPNet initialize the latent without utilizing measurement information and therefore lack structural guidance from the measurement. Consequently, they often produce reconstructions with distorted object structures, particularly in the denoising and compressed sensing results. RF-Inversion incorporates measurement information during initialization and thus restores the overall image structure more reliably than random and NPNet. However, because the latent is obtained by directly inverting the measurement, it can deviate from the Gaussian prior, making it more difficult to generate realistic images. As a result, in tasks such as super-resolution and deblurring, RF-Inversion often fails to accurately recover fine high-frequency details. In contrast, RA-DGI leverages both retrieval-based initialization and dual-guided inversion, enabling it to preserve global structure while recovering realistic high-frequency details.

Figure 3. Qualitative comparison on four inverse problems. The displayed examples illustrate improved structural fidelity and detail with RA-DGI
Inference-Time Scaling for Noise Initialization
Our method selects the initial noise through an inversion procedure, which requires approximately twice the computational time of standard random initialization. Recent work has shown that inference-time scaling can improve diffusion model performance by allocating additional compute at test time [21]. Motivated by this idea, we investigate whether similar scaling strategies can improve the performance of random noise initialization for inverse problems. We adopt two inference-time scaling strategies, random search and zero-order search, to explore multiple noise candidates during inference.
For random search, we scale the number of initial candidates in powers of two, using 2, 4, 8, 16, and 32 candidates. For zero-order search, we use two pivots and two candidates per pivot, and increase the number of optimization iterations from 1 to 5. Since ground-truth images are unavailable at test time, reference-based metrics such as PSNR, SSIM, and LPIPS cannot be used as verification signals. Instead, we employ no-reference image quality assessment metrics, specifically NIQE [23] and CLIPIQA [32], as surrogate verifiers.
We conduct experiments on the AFHQ dataset using compressed sensing as the inverse problem. Figure 4 compares our method with random initialization combined with inference-time scaling. The figure reports PSNR, SSIM, and LPIPS results under different scaling strategies, where the legend indicates the scaling algorithm (random search or zero-order search) followed by the verifier used (NIQE or CLIPIQA). When the same amount of inference-time compute is allocated, our method consistently achieves better reconstruction performance across all metrics. Furthermore, for PSNR and LPIPS, increasing the computational budget does not lead to noticeable improvements when inference-time scaling is applied to random initialization. This is likely because the verifiers rely on no-reference image quality metrics rather than ground-truth supervision. Although certain combinations achieve slightly higher SSIM than our method, the additional computational cost required makes our approach more efficient.
Figure 4. Effect of inference-time scaling compared with our initialization strategy. The x-axis shows inference-time compute (NFE), and the y-axis shows reconstruction metrics (PSNR, SSIM, LPIPS). Although some search configurations achieve higher SSIM, RA-DGI offers a more favorable tradeoff between reconstruction quality and NFE, achieving better PSNR and LPIPS with substantially fewer function evaluations.
Ablation Studies
The proposed RA-DGI consists of three main components: retrieval-augmented initialization, prior guidance, and measurement guidance. To analyze the contribution of each component, we conduct an ablation study by removing one component at a time and comparing the resulting performance. The experiments are performed on the AFHQ dataset and evaluated on two inverse problems: super-resolution and deblurring. The results are reported in Table 2.
| Method | Retrieval | Prior | Meas. | Super-Resolution PSNR | SSIM | LPIPS | FID | Deblurring PSNR | SSIM | LPIPS | FID |
| w/o retrieval | 23.37 | 0.492 | 0.313 | 31.11 | 22.62 | 0.439 | 0.299 | 31.06 | |||
| w/o prior guidance | 24.52 | 0.580 | 0.275 | <u>21.98</u> | 24.13 | 0.580 | 0.314 | <u>25.95</u> | |||
| w/o measurement guidance | 23.59 | 0.473 | 0.231 | 23.63 | 23.22 | 0.470 | 0.261 | 26.03 | |||
| RA-DGI (Ours) | <u>24.02</u> | <u>0.534</u> | <u>0.236</u> | 21.03 | <u>23.57</u> | <u>0.525</u> | <u>0.267</u> | 23.94 |
Table 2. Ablation study of the proposed components. We analyze the contributions of retrieval-augmented initialization, prior guidance, and measurement guidance. Best and second-best results are highlighted in bold and <u>underline</u>, respectively.
In the setting without retrieval, the degraded measurement is used directly as the starting point for inversion, while the same prior- and measurement-guided inversion procedure is retained. The performance degradation observed on both tasks demonstrates that the retrieved proxy provides a more informative starting point than direct inversion of the measurement.
When prior guidance is removed, the influence of measurement guidance becomes stronger. As a result, pixel-level accuracy metrics such as PSNR and SSIM increase. However, the encoded latent deviates further from the Gaussian prior, leading to a degradation in perceptual quality as measured by LPIPS. This observation indicates that the generative prior plays an important role in maintaining realistic image structures during the reconstruction process. Conversely, when measurement guidance is removed, the encoded latent remains closer to the Gaussian prior. This improves perceptual quality, resulting in better LPIPS scores. However, because measurement information is not sufficiently encoded, the pixel-level accuracy decreases. This result underscores the importance of enforcing measurement consistency during inversion for accurate reconstruction. Therefore, both guidance mechanisms are necessary to balance pixel-level accuracy and perceptual quality. Overall, the full RA-DGI model achieves the most balanced performance across both tasks, indicating that the two guidance
terms complement each other and jointly improve reconstruction quality.
Conclusion
In this paper, we investigate initialization strategies for flow-based inverse problem solvers, a topic that remains underexplored in prior work. To address this issue, we proposed RA-DGI, a method for constructing improved initial noise for flow-based inverse solvers. Our approach first retrieves a proxy clean image from a dataset, providing a more suitable starting point for inversion. During inversion, we further introduce dual guidance that jointly enforces measurement consistency and prior regularization, ensuring that the latent trajectory remains both measurement-aware and prior-aligned. Extensive experiments on multiple inverse problems demonstrate that RA-DGI outperforms standard random initialization and existing initialization strategies in overall reconstruction performance. Furthermore, we show that improving initialization is more effective and computationally efficient than inference-time scaling strategies. Overall, our findings highlight initialization as a critical factor in flow-based inverse problem solving. We hope this work encourages further research into initialization strategies for inverse problem solvers.
Acknowlegement
This research was supported by Brain Korea 21 FOUR funded by the Ministry of Education of Korea and National Research Foundation of Korea, the Ministry of Science and ICT (MSIT) in Korea under the ITRC support program supervised by the Institute for Information Communication Technology Planning and Evaluation (RS-2020-II201749), and the National Research Foundation of Korea grant funded by the Korea government (RS-2022-00144190).
Analysis on Retrieval
We conduct additional analyses of the retrieval module using 100 FFHQ [16] images on the super-resolution task. The results are summarized in Table 3.
| Size | Encoder | Dataset | PSNR↑ | SSIM↑ | LPIPS↓ |
| 100 | 24.27(0.17) | 0.564(0.018) | 0.206(0.005) | ||
| 1000 | DINOv3 | Unsplash | 24.24(0.09) | 0.562(0.006) | 0.202(0.006) |
| 10000 | 24.39(0.06) | 0.572(0.002) | 0.198(0.001) | ||
| 1000 | DINOv3 | Unsplash | 24.28 | 0.571 | 0.195 |
| VAE | 23.84 | 0.505 | 0.205 | ||
| None | 23.71 | 0.486 | 0.231 | ||
| 1000 | DINOv3 | FFHQ | 24.05 | 0.555 | 0.191 |
| AFHQ | 24.32 | 0.552 | 0.198 | ||
| LHQ | 23.84 | 0.513 | 0.200 |
Table 3. Retrieval analysis results
Effect of retrieval dataset size We first investigate the impact of the retrieval pool size by varying the number of candidate images used for retrieval. The results are reported in the first block of Table 3. Each experiment is repeated five times, and we report the mean performance together with the standard deviation in parentheses. Overall, the reconstruction quality remains relatively stable as the retrieval pool size increases from 100 to 10,000 images. These results suggest that a large candidate pool is not essential for obtaining effective proxy images. However, when the retrieval pool contains only 100 images, we observe noticeably larger performance variance across repeated runs. This indicates that a small retrieval pool may occasionally fail to provide sufficiently relevant proxy images, leading to a less stable initialization. Considering both performance and computational efficiency, we use a retrieval pool of 1,000 images as the default setting in all experiments.
Effect of retrieval encoder We next evaluate different feature representations for image retrieval, and report the results in the second block of Table 3. Specifically, we compare DINOv3 [29] features, VAE latent features from Stable Diffusion 3 (SD3) [9], and direct retrieval based on measurement similarity. Among all variants, DINOv3 consistently achieves the best reconstruction performance. Replacing DINOv3 with the SD3 VAE encoder leads to noticeable degradation across all evaluation metrics. We attribute this behavior to the different objectives of the two encoders. DINOv3 is trained to capture high-level semantic information and semantic similarity between images, whereas the VAE encoder is primarily optimized for image compression and reconstruction. Consequently, DINOv3 produces more semantically relevant proxy images for initialization. Direct retrieval based on measurement similarity yields the worst performance, indicating that low-level measurement similarity alone is insufficient for identifying semantically meaningful proxy images. These results highlight the importance of semantic-aware retrieval for obtaining effective initialization points. The observed reconstruction metrics are consistent with the motivation for semantic features, but do not directly measure semantic relevance. Direct measurement-similarity retrieval performs worst among the tested representations in this setting; this does not exclude effective low-level retrieval strategies in other settings.
Effect of domain gap Finally, we analyze the influence of the retrieval dataset domain. The results are reported in the third block of Table 3. We construct retrieval pools using FFHQ (human face dataset), AFHQ [4] (animal face dataset), and LHQ [30] (landscape dataset), which represent progressively larger domain gaps relative to the target FFHQ test set. As expected, retrieval performance degrades as the domain gap increases. Using FFHQ as the retrieval pool yields the best performance since the retrieved proxy images are drawn from the same domain as the target images. When AFHQ is used, the retrieved images still share certain structural characteristics, such as face-centered compositions, resulting in only a modest performance drop. In contrast, LHQ consists primarily of landscape images and exhibits a substantially larger semantic gap from FFHQ. Consequently, the retrieved proxy images become less relevant to the target measurements, resulting in the largest degradation in reconstruction quality. These observations suggest that retrieval quality is closely related to semantic proximity between the retrieval dataset and the target domain.
Additional Ablation Study
Retrieval method To isolate the contribution of semantic retrieval, we compare the nearest-neighbor retrieval used in RA-DGI with random shuffling, where a proxy image is randomly sampled from the dataset. We conduct this comparison on the super-resolution task using 100 FFHQ images. The results are reported in Table 4. Nearest neighbor
| PSNR↑ | SSIM↑ | LPIPS↓ | |
| Shuffled | 23.73 | 0.509 | 0.202 |
| Nearest neighbor | 24.28 | 0.571 | 0.195 |
Table 4. Ablation of the retrieval strategy.
| Inversion method | PSNR↑ | SSIM↑ | LPIPS↓ |
| Ordinary inversion | 24.38 | 0.559 | 0.233 |
| Dual-guided inversion | 24.28 | 0.571 | 0.195 |
caption: tbl:5. Ablation of the inversion method.
| Method | PSNR↑ | SSIM↑ | LPIPS↓ |
| Random (Solve 30) | 24.55 | 0.618 | 0.182 |
| RA-DGI (Inv. 10 + Solve 20) | 24.56 | 0.621 | 0.178 |
caption: tbl:6. Compute-matched comparison.
retrieval consistently outperforms random shuffling, indicating that selecting a semantically similar proxy image is important for constructing an effective initialization. This observation is also consistent with the domain gap analysis in Sec. A, further supporting the use of semantically relevant proxy images.
Inversion method. To evaluate the effectiveness of the proposed inversion strategy, we compare the dual-guided inversion used in RA-DGI with ordinary inversion on the super-resolution task using 100 FFHQ images. The results are reported in Tab. 5. Dual-guided inversion achieves better performance on most metrics. Consistent with the ablation results in Table 2, these results indicate that the proposed inversion strategy is important for achieving balanced results in terms of pixel-level fidelity and perceptual quality.
Additional Baseline
We compare RA-DGI with random initialization under the same nominal budget of 30 NFEs: 30 solve evaluations for Random versus 10 inversion plus 20 solve evaluations for RA-DGI. The experiment is conducted on 100 FFHQ images for the super-resolution task, and the results are reported in Tab. 6. RA-DGI achieves slightly better reconstruction performance than random initialization, demonstrating its effectiveness under an equal computational budget.
Measurement Consistency and Retrieval-Induced Bias
Measurement consistency To evaluate how well the reconstructed images align with the measurements, we measure measurement consistency using , where denotes the reconstructed image. We compute this metric for one seed from the experiments reported in Table 1, with the results shown in Tab. 7. RF-Inversion [28] and RA-DGI achieve better measurement consistency than the other initialization strategies, as both incorporate the mea-
| Dataset | Task | Random | NPNet | RF-Inversion | RA-DGI |
| AFHQ | SR | 0.0055 | 0.0058 | 0.0049 | 0.0039 |
| AFHQ | Deblur | 0.0056 | 0.0059 | 0.0055 | 0.0036 |
| AFHQ | Denoise | 0.0273 | 0.0270 | 0.0266 | 0.0220 |
| AFHQ | CS | 0.0317 | 0.0411 | 0.0177 | 0.0154 |
| FFHQ | SR | 0.0080 | 0.0105 | 0.0067 | 0.0044 |
| FFHQ | Deblur | 0.0081 | 0.0097 | 0.0074 | 0.0043 |
| FFHQ | Denoise | 0.0275 | 0.0286 | 0.0304 | 0.0214 |
| FFHQ | CS | 0.0313 | 0.0379 | 0.0221 | 0.0169 |
| DIV2K | SR | 0.0107 | 0.0111 | 0.0080 | 0.0064 |
| DIV2K | Deblur | 0.0116 | 0.0119 | 0.0092 | 0.0070 |
| DIV2K | Denoise | 0.0527 | 0.0534 | 0.0476 | 0.0457 |
| DIV2K | CS | 0.0596 | 0.0820 | 0.0432 | 0.0424 |
caption: tbl:7. Measurement consistency results.
surement when constructing the initial noise through inversion. RA-DGI further improves measurement consistency over RF-Inversion by explicitly incorporating measurement guidance during the inversion process.
Retrieval-induced bias We qualitatively examine the effect of incorrect retrieval on the final reconstruction. Figure 5 shows cases where the retrieved images are semantically unrelated to the ground truth. Despite this mismatch, the reconstructions remain consistent with the measurements and successfully recover the target identity and structure without visibly inheriting the semantics or textures of the retrieved images. We attribute this robustness to the measurement-consistency guidance applied during both inversion and sampling, which continuously anchors the reconstruction trajectory to the observation and limits the influence of irrelevant content from the retrieved image.

Figure 5. Retrieval-failure cases.
Wall-Clock Efficiency
We report the per-image wall-clock latency averaged over 100 images in Table 8. Except for RA-DGI, the compared methods exhibit similar runtimes. Although RF-Inversion [28] additionally performs inversion, it relies on ordinary inversion without gradient computation and therefore incurs only a small computational overhead. In contrast, RA-DGI requires an additional inversion process whose runtime is comparable to that of sampling, resulting in approximately twice the latency of random initialization. The additional operations required by RA-DGI, such as retrieval and nearest-neighbor search, are computationally inexpensive and have a negligible impact on the overall runtime. Therefore, we expect the relative trends observed in the NFE-based comparison in Figure 4 to remain largely consistent when evaluated in terms of wall-clock time.
| Method | Time/image (s) |
| Random | 5.9 |
| NPNet | 6.1 |
| RF-Inversion | 6.3 |
| RA-DGI | 11.2 |
Table 8. Per-image wall-clock latency.
Compatibility with FLAIR
To evaluate the compatibility of RA-DGI with existing inverse problem solvers, we integrate its initialization strategy with FLAIR [8] and evaluate the resulting method on the super-resolution task using 100 FFHQ images. As shown in Table 9, incorporating RA-DGI consistently improves reconstruction performance over the original FLAIR initialization. These results demonstrate that RA-DGI can be readily combined with different inverse problem solvers to further improve their performance.
| Method | PSNR | SSIM | LPIPS |
| FLAIR | 23.99 | 0.596 | 0.538 |
| FLAIR + RA-DGI | 24.03 | 0.600 | 0.532 |
Table 9. Compatibility with FLAIR.
Theoretical Analysis
The following theorem characterizes the probability that a random Gaussian latent lies within a given distance of a fixed target latent. This result provides a quantitative explanation for why a randomly sampled initialization is extremely unlikely to be close to the latent corresponding to a particular image.
Theorem 1 (Distribution of the Distance to a Fixed Latent). Let be a standard Gaussian latent in , and let be a fixed deterministic latent (e.g., the ground truth inversion latent). For any radius , define
Then follows a non-central chi-square distribution with degrees of freedom and non-centrality parameter , i.e.,
Consequently, the probability that a random Gaussian latent lies within an ball of radius centered at is
where denotes the cumulative distribution function of the non-central chi-square distribution.
Proof. Let and let be a fixed deterministic vector. Define
Since independently, each shifted variable
Therefore, is the square of a Gaussian random variable with mean and unit variance. By definition, the sum of squares of independent Gaussian variables with means and unit variances follows a non-central chi-square distribution with degrees of freedom and non-centrality parameter
In our case , hence
Consequently,
where denotes the non-central chi-square distribution with degrees of freedom and non-centrality parameter . Finally,
which completes the proof. □
Theorem 1 shows that the distance between a random Gaussian latent and a fixed target latent follows a non-central chi-square distribution. Therefore, the probability that a random initialization falls within any prescribed neighborhood of the target latent can be computed directly from the corresponding cumulative distribution function.
Remark 1 (Typical norm of Gaussian latents). For , define . Then follows a chi-square distribution with degrees of freedom, i.e., , with mean and variance . Although the standard deviation of grows as , its mean grows as . Therefore, its relative standard deviation, , decreases as the dimension increases. Consequently, deviations of from its mean become small relative to , supporting the approximation for a typical Gaussian latent in high dimensions.
Numerical Example
We use Theorem 1 to illustrate the proximity advantage of RA-DGI initialization over random Gaussian initialization. Let denote the latent obtained by inverting the ground truth image, and let denote a randomly sampled Gaussian latent. Theorem 1 characterizes the distribution of their squared Euclidean distance, allowing us to compute the probability that lies within a given radius of .
In our experiments, the mean Euclidean distance be tween the ground truth latent and the RA-DGI initialization is approximately 516 across 100 evaluation samples. We therefore set as a representative radius in Theorem 1. To specify the non-centrality parameter, we assume that sufficiently accurate inversion maps ground truth images approximately to the model’s standard Gaussian prior, and thus regard as a typical Gaussian latent. The SD3 [9] latent space has dimensionality . At this dimensionality, the norm concentration described in Remark 1 motivates the approximation .
Using Theorem 1 with these parameters, the probability that a randomly sampled Gaussian latent lies within this representative distance of the target is
This result shows that a random Gaussian initialization is highly unlikely to match RA-DGI’s average proximity to the target latent, supporting RA-DGI’s proximity advantage.
Limitation
Our method focuses on improving the initial noise rather than modifying the inverse problem solver itself. In this work, experiments are conducted using the FlowDPS [18] solver. Therefore, the overall reconstruction performance is inherently bounded by the capability of this solver. Previous studies have shown that FlowDPS performs well on tasks that preserve sufficient structural information, such as super-resolution and deblurring. However, it struggles with problems where the measurement provides extremely limited information, such as inpainting and colorization. Figure 6 illustrates such cases. In the inpainting example, even with the proposed initialization, the missing region still contains inaccurate content. Similarly, in the colorization example, both random initialization and the proposed method struggle to produce colorized results that faithfully match the ground truth. These results indicate that improving initialization alone cannot fully overcome the solver’s intrinsic limitations when the measurement is highly ambiguous.

Figure 6. Failure cases on tasks where the underlying solver is weak. Even with the proposed initialization, it remains difficult to overcome the inherent limitations of the original solver.
Additional Results
We provide additional qualitative comparisons across three datasets. Figure 7, Figure 8, and Figure 9 present representative results on the AFHQ, FFHQ, and DIV2K datasets, respectively. Across different datasets and degradation settings, the proposed initialization consistently produces reconstructions that are visually closer to the ground truth, while reducing structural distortions and unrealistic artifacts.

Figure 7. Additional qualitative comparisons on the AFHQ dataset.

Figure 8. Additional qualitative comparisons on the FFHQ dataset.

Figure 9. Additional qualitative comparisons on the DIV2K dataset.
References
- [1]Eirikur Agustsson and Radu Timofte. NTIRE 2017 challenge on single image super-resolution: Dataset and study. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017. 5
- [2]Cagan Alkan, Julio Oscanoa, Daniel Abraham, Mengze Gao, Aizada Nurdinova, Kawin Setsompopp, John M Pauly, Morteza Mardani, and Shreyas Vasanawala. Variational diffusion models for blind mri inverse problems. In NeurIPS 2023 workshop on deep learning and inverse problems, page 2, 2023. 2DOI
- [3]Heli Ben-Hamu, Omri Puny, Itai Gat, Brian Karrer, Uriel Singer, and Yaron Lipman. D-flow: Differentiating through flows for controlled generation. In Proceedings of the 41st International Conference on Machine Learning, pages 3462–3483. PMLR, 2024. 2
- [4]Yunjey Choi, Youngjung Uh, Jaejun Yoo, and Jung-Woo Ha. StarGAN v2: Diverse image synthesis for multiple domains. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8188–8197, 2020. 5, 11
- [5]Hyungjin Chung, Jeongsol Kim, Michael Thompson Mc-cann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In The Eleventh International Conference on Learning Representations, 2023. 1, 2arxiv.org/abs/2209.14687
- [6]Giannis Daras, Hyungjin Chung, Chieh-Hsin Lai, Yuki Mitsufuji, Jong Chul Ye, Peyman Milanfar, Alexandros G Dimakis, and Mauricio Delbracio. A survey on diffusion models for inverse problems. arXiv preprint arXiv:2410.00083, 2024. 2arxiv.org/abs/2410.00083
- [7]Alexander Denker, Moshe Eliasof, Zeljko Kereta, and Carola-Bibiane Schönlieb. Trajectory stitching for solving inverse problems with flow-based models. arXiv preprint arXiv:2602.08538, 2026. 2arxiv.org/abs/2602.08538
- [8]Julius Erbach, Dominik Narnhofer, Andreas Robert Dombos, Bernt Schiele, Jan Eric Lenssen, and Konrad Schindler. Solving inverse problems with FLAIR. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. 1, 2, 13
- [9]Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first international conference on machine learning, 2024. 1, 11, 15arxiv.org/abs/2403.03206
- [10]Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. ReNO: Enhancing one-step text-to-image models through reward-based noise optimization. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2arxiv.org/abs/2406.04312
- [11]Luca Eyring, Shyamgopal Karthik, Alexey Dosovitskiy, Nataniel Ruiz, and Zeynep Akata. Noise hypernetworks: Amortizing test-time compute in diffusion models. In The Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025. 2
- [12]Berthy T Feng, Jamie Smith, Michael Rubinstein, Hui-wen Chang, Katherine L Bouman, and William T Freeman. Score-based diffusion models as principled priors for inverse imaging. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10520–10531, 2023. 2arxiv.org/abs/2304.11751
- [13]Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9380–9389, 2024. 2arxiv.org/abs/2404.04650
- [14]Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 5
- [15]Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. In NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications, 2021. 5arxiv.org/abs/2207.12598
- [16]Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401–4410, 2019. 5, 11DOI
- [17]Bahjat Kawar, Michael Elad, Stefano Ermon, and Jiaming Song. Denoising diffusion restoration models. In Advances in Neural Information Processing Systems, 2022. 1, 2arxiv.org/abs/2201.11793
- [18]Jeongsol Kim, Bryan Sangwoo Kim, and Jong Chul Ye. Flowdps: Flow-driven posterior sampling for inverse problems. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12328–12337, 2025. 1, 2, 3, 4, 5, 15arxiv.org/abs/2503.08136
- [19]Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, 2023. 1, 3arxiv.org/abs/2210.02747
- [20]Xingchao Liu, Chengyue Gong, and qiang liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In The Eleventh International Conference on Learning Representations, 2023. 1, 3arxiv.org/abs/2209.03003
- [21]Nanye Ma, Shangyuan Tong, Haolin Jia, Hexiang Hu, Yu-Chuan Su, Mingda Zhang, Xuan Yang, Yandong Li, Tommi Jaakkola, Xuhui Jia, et al. Scaling inference time compute for diffusion models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 2523–2534, 2025. 2, 6DOI
- [22]Morteza Mardani, Jiaming Song, Jan Kautz, and Arash Vahdat. A variational perspective on solving inverse problems with diffusion models. In The Twelfth International Conference on Learning Representations, 2024. 2arxiv.org/abs/2305.04391
- [23]Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a “completely blind” image quality analyzer. IEEE Signal processing letters, 20(3):209–212, 2012.DOI
- [24]Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024.arxiv.org/abs/2307.01952
- [25]Mehrsa Pourya, Bassam El Rawas, and Michael Unser. FLOWER: A flow-matching solver for inverse problems. In The Fourteenth International Conference on Learning Representations, 2026.arxiv.org/abs/2509.26287
- [26]Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2):3, 2022.arxiv.org/abs/2204.06125
- [27]Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022.arxiv.org/abs/2112.10752
- [28]Litu Rout, Yujia Zhang, Nathaniel Ruiz, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Semantic image inversion and editing using rectified stochastic differential equations. In The Thirteenth International Conference on Learning Representations, 2025.arxiv.org/abs/2410.10792
- [29]Oriane Siméoni, Huy V Vo, Maximilian Seitzer, Federico Baldassarre, Maxime Oquab, Cijo Jose, Vasil Khalidov, Marc Szafraniec, Seungeun Yoon, Michaël Ramamonjisoa, et al. Dinov3. arXiv preprint arXiv:2508.10104, 2025.arxiv.org/abs/2508.10104
- [30]Ivan Skorokhodov, Grigorii Sotnikov, and Mohamed Elho-seiny. Aligning latent and image spaces to connect the un-connectable. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14144–14153, 2021.DOI
- [31]Unsplash. Unsplash dataset. https://unsplash.com/data, 2020.
- [32]Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI conference on artificial intelligence, pages 2555–2563, 2023.arxiv.org/abs/2207.12396
- [33]Yinhuai Wang, Jiwen Yu, and Jian Zhang. Zero-shot image restoration using denoising diffusion null-space model. In The Eleventh International Conference on Learning Representations, 2023.arxiv.org/abs/2212.00490
- [34]Chenfei Wu, Jiahao Li, Jingren Zhou, Junyang Lin, Kaiyuan Gao, Kun Yan, Sheng-ming Yin, Shuai Bai, Xiao Yu, Yilei Chen, et al. Qwen-image technical report. arXiv preprint arXiv:2508.02324, 2025.arxiv.org/abs/2508.02324
- [35]Rongyuan Wu, Tao Yang, Lingchen Sun, Zhenqiang Zhang, Shuai Li, and Lei Zhang. Towards semantics-aware real-world image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 25456–25467, 2024.arxiv.org/abs/2311.16518
- [36]Katherine Xu, Lingzhi Zhang, and Jianbo Shi. Good seed makes a good crop: Discovering secret seeds in text-to-image diffusion models. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3024–3034. IEEE, 2025.arxiv.org/abs/2405.14828
- [37]Bingliang Zhang, Wenda Chu, Julius Berner, Chenlin Meng, Anima Anandkumar, and Yang Song. Improving diffusion inverse problem solving with decoupled noise annealing. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 20895–20905, 2025.DOI
- [38]Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586–595, 2018.arxiv.org/abs/1801.03924
- [39]Zikai Zhou, Shitong Shao, Lichen Bai, Shufei Zhang, Zhiqiang Xu, Bo Han, and Zeke Xie. Golden noise for diffusion models: A learning framework. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17688–17697, 2025.

