acceptodds
Withdrawn from ICLR 2027

RealRestorer: Towards Generalizable Real-World Image Restoration with Large-Scale Image Editing Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Image restoration under real-world degradations is critical for downstream tasks such as autonomous driving and object detection. However, existing restoration models are often limited by the scale and distribution of their training data, resulting in poor generalization to real-world scenarios. Recently, large-scale image editing models have shown strong generalization ability in restoration tasks, especially for closed-source models like Nano Banana Pro, which can restore images while preserving consistency. Nevertheless, achieving such performance with those large universal models requires substantial data and computational costs. To address this issue, we construct a large-scale dataset covering nine common real-world degradation types and train a state-of-the-art open-source model to narrow the gap with closed-source alternatives. Furthermore, we introduce RealIR-Bench, which contains 464 real-world degraded images and tailored evaluation metrics focusing on degradation removal and consistency preservation. Extensive experiments demonstrate our model ranks first among open-source methods, achieving state-of-the-art performance.

Primary area:
datasets and benchmarks
Keywords:
Image Restoration

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_i76fOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a high-quality large-scale dataset (RealIR-1M) and a benchmark (RealIR-Bench) for general-purpose image restoration. The dataset covers 9 common real-world degradation types (blur, noise, low-light, etc.) and contains a substantial number of samples (~1.65M). With this dataset, the paper proposes a real-world image restoration model, RealRestorer, by fine-tuning Step1X-Edit.

Critical strengths

1.

The engineering contribution looks solid and is described in enough detail to follow: a per-degradation synthesis pipeline, CLIP retrieval prompts, and the full training configuration (optimizer, schedule, steps, batch, hardware, etc.), with a plan to release the model and the pipeline. The high-quality large-scale dataset for general-purpose image restoration is also a main strength. If the dataset and the benchmark are also released, I think they could be useful to the community by themselves.

Critical weaknesses

1.

First of all, the text defines LPS = 1 - LPIPS (higher is better), but I think that the "LPS" columns in Tables 1 and 2 appear to indicate LPIPS itself (with a down arrow), and the Table 2 caption calls the same column "LPIPS". Since I was confused interpreting the table, please use notations consistently.

As far as I understand, the final score FS is a product of two terms: FS = 0.2 x (1 - LPIPS) x RS, where the (1 - LPIPS) measures how much the output stays consistent with the input and the RS measures how much degradation is removed. On the "1 - LPIPS" term, the proposed model is the worst among all open-source models: its average "1-LPIPS" is 1-0.445=0.555, while the other open-source models range from 0.675(=1-0.325) to 0.798(=1-0.202) (Table 2, Avg Total).

That is, I think that its first-ranked FS comes from the removal term instead of consistency term "1 - LPIPS": its RS of 1.312 is much higher than the next-ranked open-source model Qwen (0.931), which is large enough to outweigh the worst consistency in the product ( {(1.312-0.931)/0.931}x100 ~ 41% higher in RS ). In other words, the model seems to win by changing the image the most, not by preserving it, while the abstract presents consistency preservation as a central property.

Thus, I think there is an internal tension that the text does not discuss and this is the main reason for my recommendation.

2.

Second, the state-of-the-art claim does not seem matched by the comparison set, and a key part of the data pipeline is not specified. As far as I checked, every quantitative table (Tables 1, 2, 3, and 8) compares only against image editing models. That is, no dedicated restoration or all-in-one restoration model appears in any table, even on FoundIR (all-in-one restore), Flare7K++ (Deflare), and UHDM (Demoiré) (Tables 3 and 8), where such models are the established reference points. Small restoration models (EnlightenGAN, Retinexformer) appear only in a qualitative figure (Figure 2).

That is, I think "achieving state-of-the-art performance" (Abstract) is supported only within the editing-model family, and even there the model is second on average PSNR on the public benchmarks (20.88 vs. 22.21; Table 8). In addition, the clean references for the real-world training pairs are generated by "appropriate editing models" (Appendix A.2), but their identity is not given anywhere.

Thus, in my opinion, the paper could not support its central claims until the comparison model set is extended or the claims are narrowed. Also, please specify which models generated the clean references.

Questions for authors
  1. Section 3.2 says that "the intermediate checkpoint with the best generalization capability" is selected, and Figure 6 plots the performance against training steps on RealIR-Bench. Was this curve measured on the RealIR-Bench evaluation set itself, or on a separate validation set? If the former, I think the reported checkpoint could have been selected on the same benchmark used for Tables 1 and 2, which is a selection step the baseline editing models were not given. Could you state the exact selection rule and the data it was computed on?
  2. As far as I understand, both RealIR-Bench and the real-world training data are collected from the internet (Sec. 3.1, 4.1). Was any deduplication performed between the two? Could you report the near-duplicate rate if possible?
  3. The benchmark size is 464 in the abstract and Sec. 4.1 but 456 in Table 7. And the training-sample counts are different across Sec. 4.4 (about 1M / 100K), Table 5 (1,567,492 / 87,585), and Table 6 (1.5M / 80K). Could you please write these numbers accurately and consistently?
  4. I was confused by several citations that appear to come from unrelated fields: Benz (2017) is a solar-physics review but cited for lens flare, Tong & White (1996) and Al-Salem et al. (2015) are about plastics degradation. Could you clarify or correct these references? And please check all the references in the paper once more, so that every citation actually points to a work in the related field.

Minor issues:

  • Typos and formatting: "Apeendix F" (Sec. 4.4), "removalremoval" (Table 8 caption), and a broken sentence ("...consistency. which means...") in the Figure 7 caption.
Reviewer_rswaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

RealRestorer is an open-source image restoration framework fine-tuned from Step-1X.Edit, designed to generalize across nine common real-world degradations. To bridge the gap between synthetic and authentic degradations, the authors propose RealIR-1M, a dataset comprising ~100K strictly filtered real-world degradation pairs alongside 1.5M synthetic samples. They further propose a two-stage progressively-mixed training strategy with a DiT backbone: Stage 1 leverages large-scale synthetic data for transfer learning, while Stage 2 fine-tunes on a mixture of 80% real and 20% synthetic data. Additionally, the authors introduce RealIR-Bench, a non-reference benchmark of 464 authentic images, along with a VLM-based evaluation framework that computes a Final Score (FS) by balancing restoration quality (RS) and content fidelity (LPS).

Critical strengths

Strength:

1: Scalability Potential — This work demonstrates the scalability potential of leveraging open-source edit models under non-real-captured paired data settings. Previous approaches typically required real-world photography to construct degradation pairs for each specific scenario, making data scaling extremely difficult. In contrast, this work's data pipeline requires at least one side (either the degraded image or the ground truth) to be obtained through synthesis with filtering, significantly lowering the barrier for data scaling.

2: Engineering Reference Value — This work provides engineering reference value by successfully pushing an open-source edit model (Step-1X.Edit) to approach the performance of closed-source counterparts, offering a reproducible blueprint for the community to uplift their own open-source models.

Critical weaknesses
  1. Insufficient details on dataset construction. Since the dataset and its construction pipeline are claimed as major contributions, the collection, generation, and filtering procedures should be described more explicitly. For the synthetic data, the specific sources and number of unique clean images, as well as the models and criteria used for quality and watermark filtering, remain unclear. Section 3.1 also states that SAM-2 and MiDaS are used to filter unsuitable samples and provide structural cues, but their filtering criteria and precise roles in degradation synthesis are not sufficiently specified. More importantly, for the real-world degradation pairs, the paper does not identify the models and versions used to generate pseudo-clean targets, their assignment to different degradation categories, or the generation prompts. It is also unclear whether prompts were adjusted per model or image, and whether multiple candidates were generated and selected. These details are essential for reproducing and assessing the proposed dataset.

  2. Insufficient controlled experiments to establish the data pipeline’s contribution. The paper compares RealRestorer with the original Step1X-Edit and evaluates different combinations of its own training data, but does not compare the proposed data against existing restoration training datasets using the same backbone and a matched training budget. Evaluation on the FoundIR test set does not substitute for this comparison. For example, starting from the same Stage 1 checkpoint and replacing the proposed Stage 2 real-to-pseudo-clean pairs with existing paired restoration data, while controlling task coverage, data volume, and optimization settings, would help isolate the contribution of the proposed data. In addition, the training pipeline is validated only on Step1X-Edit, leaving its dependence on the backbone’s pretrained priors and its applicability to other editing models insufficiently explored.

  3. Limited restoration baselines and insufficient analysis of prompt sensitivity. The quantitative comparisons on RealIR-Bench, FoundIR, primarily include general-purpose image editing models. Competitive task-specific restoration models and all-in-one restoration methods are missing. Since RealIR-Bench evaluates individual degradation categories with explicit task-specific instructions, including corresponding specialist models would provide a more informative assessment of restoration performance, particularly content fidelity. Furthermore, although the fixed evaluation prompts improve reproducibility, they may not be equally suitable for all editing models. A prompt-sensitivity study, or model-specific prompt selection on a separate validation set, would help establish whether the reported rankings are robust. The current evidence primarily supports competitiveness among the evaluated editing models rather than a broader state-of-the-art claim for image restoration.

  4. Remaining concerns about content fidelity. Although the paper includes fidelity-related metrics and a user study, some qualitative examples exhibit changes in illumination, color, and local appearance. Further analysis is needed to distinguish necessary degradation removal from unnecessary modification of the original scene. Both PSNR and SSIM decrease for compression restoration. These results suggest that improved restoration strength does not consistently translate into better fidelity. Additional failure cases and task-level analysis would help clarify when the model preserves the observed content and when it introduces excessive enhancement or unsupported detail.

Questions for authors

Please refer to above weakness.

Reviewer_aBTrOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses the limited generalization ability of existing image restoration models in complex, real-world degradation scenarios by exploring the generative priors of large-scale image editing models. To this end, the authors construct RealIR-1M, a large-scale dataset covering nine common degradation types that combines synthetic data with curated real-world degradation pairs, and fine-tune an open-source image editing model using a two-stage progressively-mixed training strategy to develop the RealRestorer model. Additionally, the paper introduces RealIR-Bench, an evaluation benchmark consisting of 464 real-world non-reference degraded images, alongside a composite evaluation metric that combines Vision-Language Model (VLM) scoring with perceptual similarity (LPIPS). Experimental results indicate that RealRestorer outperforms the evaluated open-source image editing baselines on various natural image restoration tasks, achieving performance levels comparable to certain closed-source commercial systems.

Critical strengths
  1. The construction of a comprehensive mixed-degradation dataset (RealIR-1M) represents a substantial and highly practical engineering effort. Rather than relying solely on simplistic synthetic corruptions, the authors developed a robust pipeline that pairs web-sourced real-world degraded images with clean references, validated through semantic filtering and edge-shift analysis to ensure visual alignment. This data-centric approach is consequential because bridging the synthetic-to-real gap is a critical bottleneck in current image restoration, making the curated dataset a valuable resource for the community dealing with natural in-the-wild image corruptions.

  2. The proposed two-stage progressively-mixed training strategy provides an effective and well-designed solution for transferring high-level generative priors to low-level restoration tasks. By first training on diverse synthetic degradations and subsequently fine-tuning on a carefully balanced mixture of real and synthetic data, the framework successfully mitigates the overfitting and structural inconsistency issues that typically arise when training solely on real-world pairs. This empirical stability demonstrates a solid engineering practice that supports the visual quality improvements seen in the natural image results.

Critical weaknesses
  1. While the comparison with large-scale image editing models (e.g., GPT-Image-1.5, Nano Banana Pro) offers interesting insights, the empirical evaluation would benefit significantly from including dedicated state-of-the-art image restoration networks (such as recent All-in-One models). Including these expert baselines would provide a much more comprehensive context for the image restoration community and help fully substantiate the state-of-the-art performance claims.

  2. The introduction of a VLM-based metric (Restoration Score) is an innovative attempt, but its robustness and alignment with human perception for complex low-level degradations might require deeper validation. Using a zero-shot VLM (Qwen3-VL-8B-Instruct) introduces potential scoring biases. Additionally, combining this score with LPIPS into a single Final Score (FS=0.2⋅LPS⋅RSFS = 0.2 \cdot LPS \cdot RS) would benefit from more rigorous theoretical or empirical justification to fully assure the community of the benchmark's reliability.

Questions for authors
  1. The evaluation in Tables 1 and 2 focuses heavily on large-scale image editing models. To better position this work within the image restoration community, could the authors explain the rationale for omitting comparisons with dedicated state-of-the-art All-in-One image restoration models? Providing a quantitative comparison with these expert baselines during the rebuttal would significantly help fully substantiate the state-of-the-art performance claims.

  2. The paper emphasizes generalizable real-world restoration, but the evaluation focuses primarily on natural photographic images. How well do these generative priors transfer to domain-specific "hard vision" scenarios (e.g., remote sensing, medical imagery) where structural hallucinations are particularly problematic? Clarification on the intended scope or preliminary results on out-of-domain data would provide valuable context.

  3. The Restoration Score relies on a zero-shot VLM (Qwen3-VL-8B-Instruct). Has this evaluator been validated for potential scoring biases when judging specific low-level degradations? Furthermore, could the authors provide deeper theoretical or empirical justification for the specific mathematical formulation of the Final Score (FS=0.2⋅LPS⋅RSFS = 0.2 \cdot LPS \cdot RS)?

  4. While RealRestorer utilizes a powerful DiT backbone (Step1X-Edit) to achieve impressive visual results, there is limited discussion regarding its computational efficiency during inference. Could the authors provide a comparison of inference time, FLOPs, and parameter counts between RealRestorer and dedicated expert restoration models? Understanding the trade-off between restoration quality and computational cost is crucial for real-world deployment.

  5. RealIR-Bench contains 464 images across nine degradation categories, averaging around 50 images per task. Considering the extreme diversity of real-world corruptions, is this sample size sufficient to comprehensively capture the long-tail distribution? I suggest briefly discussing plans to scale this benchmark or providing confidence intervals for the reported metrics to ensure statistical significance.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.