acceptodds
Under review as a conference paper at ICLR 2027

ProtoFlow: Universal One-step Flow Restorer via Prototype Guidance

Abstract

Universal image restoration seeks to recover high-quality images from diverse degradations within a unified framework. Recent diffusion- and flow-based methods achieve strong restoration performance by leveraging powerful generative priors, yet remain costly due to multi-step computation and offer only limited control through text prompts. In addition, restoration performance deteriorates when explicit degradation information is unavailable. To address these issues, we rethink image restoration as a short-horizon transport from degraded latent toward the clean-image manifold, where the timestep naturally serves as a restoration-strength controller. We further introduce ProtoFlow, a prototype-guided one-step flow model that leverages Gaussian prototypes to enhance the representation of both degradation and semantics, thereby bridging the gap between settings with and without explicit degradation. On CoTIR-Bench, ProtoFlow significantly outperforms the CoTIR-4B on full-reference metrics, remains competitive on no-reference metrics, and runs about faster. These results demonstrate that one-step flow restoration can achieve strong performance, flexible intensity control, and high efficiency when paired with structured prototype priors.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Low-level Vision, Image Restoration, Image Generation

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_VCviOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ProtoFlow adapts a pretrained FLUX.2-4B model for one-step image restoration, starting directly from the degraded image’s latent representation. It learns Gaussian prototype banks for degradation and semantic information, retrieves prototype-based guidance, and fine-tunes the restoration backbone using LoRA. The method also uses the timestep to adjust restoration strength. Experiments on CoTIR-Bench show improved full-reference metrics and substantially faster inference than CoTIR-4B.

Critical strengths
  1. Clear improvements in restoration accuracy and speed. Under vague prompts, ProtoFlow improves PSNR over CoTIR-4B by 1.82 dB and reduces inference time by approximately 7.7×. SSIM and LPIPS also improve.

  2. Useful ablations support the prototype design. The paper evaluates the degradation and semantic prototype banks separately and together. Both contribute to performance, and their combination produces the best results in the reported ablation.

  3. The method offers useful flexibility. It supports both vague and precise degradation prompts, with a smaller performance gap between these settings than CoTIR. Qualitative examples also demonstrate restoration-strength adjustment through the timestep.

Critical weaknesses
  1. Gains in no-reference quality are limited. Although ProtoFlow improves full-reference metrics and inference speed, its Q-Align and LIQE scores are slightly lower than CoTIR’s, while the improvement in CLIP-IQA+ is small. Further evidence, such as a human preference study, would help establish whether these gains translate into better visual quality.

  2. The generalization of the prototype design needs further validation. The prototype banks use predefined degradation categories and organizational structures. Existing ablations support their effectiveness on the evaluated benchmark, but it remains unclear whether they outperform prototype-free or simpler conditioning alternatives on unseen degradations. A setting without any text prompt is also not separately evaluated.

Questions for authors

As described in "Critical Weaknesses".

4 comments

We sincerely thank Reviewer VCvi for the constructive feedback. We address each concern directly below.

R-W1: Human preference study.

While ProtoFlow's Q-Align and LIQE scores are marginally below CoTIR-4B (by less than 1.5%), no-reference metrics alone cannot determine whether visually plausible restorations preserve the actual scene content or introduce hallucinations. To directly assess these aspects, we conducted a human study with 50 participants, each evaluating all compared methods on 10 independently sampled CoTIR-Bench images. The outputs were anonymized and randomized, and participants rated visual quality (VQ), fidelity to ground truth (Fid.), and absence of hallucinations (AH) on a 0–100 scale.

ProtoFlow achieves higher mean human ratings than CoTIR-4B across all three criteria under both prompt settings:

MethodVisual quality ↑Fidelity ↑Absence of hallucinations ↑
PromptIR39.7242.2486.52
OneRestore†56.0855.8097.04
InstructIR49.9650.8896.40
InstructIR†54.4856.7297.96
DA-CLIP58.4857.0088.04
HYPIR†66.8861.2073.20
LucidFlux36.6832.0835.48
CoTIR-4B76.0074.5696.96
CoTIR-4B†77.0476.2097.36
ProtoFlow79.4479.0499.84
ProtoFlow†78.5278.6099.80

† indicates that a precise degradation prompt is provided.

Under vague prompts, the proportion of visual-quality ratings ≥80 increases from 80% to 92%, while fidelity improves from 74.56 to 79.04. Under precise prompts, the proportion of high visual-quality ratings increases from 80% to 90%. These results provide direct human evidence that ProtoFlow improves perceived visual quality and reference fidelity while reducing hallucinations, complementing its superior full-reference metrics.

We have incorporated the complete human evaluation into Table 1, rating distributions into Fig. 3, and the evaluation protocol into Appendix D.

R-W2 1): Generalization to unseen degradation combinations.

We further evaluated ProtoFlow on an independent external benchmark [1] containing 200 images across eight degradation combinations absent from our training data. All methods were evaluated on the same images without additional fine-tuning.

ProtoFlow outperforms all compared methods across all six full-reference and no-reference metrics under both vague and precise prompts. More importantly, comparison with the same one-step architecture without prototypes directly establishes the contribution of prototype guidance under this distribution shift:

MethodPSNR ↑SSIM ↑LPIPS ↓CLIP-IQA+ ↑Q-Align ↑LIQE ↑
Input18.510.45210.74600.37942.18921.0677
PromptIR19.800.54170.60830.37172.19871.0747
OneRestore†21.250.64560.48700.39522.40081.2690
InstructIR19.890.58590.50740.37752.61461.2320
InstructIR†20.080.59110.50790.39322.60961.2145
DA-CLIP20.890.61760.46780.42962.85961.6523
HYPIR†20.810.62520.32260.48553.28142.2044
LucidFlux18.270.54790.42460.50893.24682.1552
CoTIR-4B21.680.69410.40070.42232.89221.7054
CoTIR-4B†21.850.69840.38350.43002.97271.7955
ProtoFlow (w/o Proto.)20.870.65320.41230.42183.05762.0476
ProtoFlow† (w/o Proto.)21.350.68750.39760.45363.23582.1563
ProtoFlow22.670.70270.28960.51253.43022.4144
ProtoFlow†22.760.70240.29020.50913.42192.4126

† indicates a precise degradation prompt. “w/o Proto.” denotes the variant without prototype representation.

Under vague prompts, prototype guidance improves PSNR by 1.80 dB and reduces LPIPS from 0.4123 to 0.2896. Under precise prompts, the PSNR improvement is 1.41 dB, with consistent gains across all six metrics.

These results demonstrate zero-shot generalization to previously unseen combinations of degradation types and show that prototype guidance remains effective beyond the original CoTIR-Bench evaluation distribution. The complete six-metric comparison and visual examples are provided in Appendix E.2 (Table 5 and Fig. 14).

[1] Zhu K, et al. An Intelligent Agentic System for Complex Image Restoration Problems. ICLR, 2025.

R-W2 2): Restoration with an empty prompt.

We additionally evaluated ProtoFlow with an empty text prompt, passing an empty string to the text encoder while retaining the same prototype conditioning as in the vague-prompt setting. We use the same trained checkpoint and CoTIR-Bench test images.

PromptPSNR ↑SSIM ↑LPIPS ↓
Empty26.320.87400.0834
Vague26.340.87410.0836

The results are nearly identical: removing the generic text instruction changes PSNR by only 0.02 dB, with negligible differences in SSIM and LPIPS. This directly demonstrates that ProtoFlow does not require a textual degradation instruction to maintain its restoration accuracy on CoTIR-Bench, as the learned prototype guidance provides effective degradation and semantic conditioning.

Thank you for the careful rebuttal which addressed some of my concerns. I will maintain my positive score.

Dear Reviewer VCvi,

Thank you for taking the time to read our response and for your positive score. We appreciate your constructive feedback and are glad the rebuttal addressed your concerns. If you have any further concerns, we would be happy to answer them.

Best regards,

The Authors

Reviewer_5x4XOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ProtoFlow, a one-step flow-based framework for universal image restoration built upon pretrained FLUX.2. The method treats restoration as a short transport from degraded to clean latents and uses the timestep to control restoration strength. It also introduces two Gaussian prototype banks to represent degradation patterns and semantic information, which are injected into the restoration backbone through lightweight conditioning. Experiments on CoTIR-Bench show improved full-reference metrics and substantially faster inference compared with CoTIR-4B.

Critical strengths
  1. The improvements in Table 1 are substantial. Compared with CoTIR-4B, ProtoFlow improves PSNR from 24.52 to 26.34 dB and reduces inference time from 4.602s to 0.594s. I think this is a meaningful practical improvement, particularly for a large generative restoration model.

  2. Table 2 shows that degradation and semantic prototypes provide complementary benefits, improving PSNR by 1.32 dB over the one-step baseline. Appendix Table 5 further shows that Gaussian prototypes outperform conventional prototypes. These results provide reasonable evidence that the prototype design contributes beyond simply using one-step inference.

Critical weaknesses
  1. I think the motivation is intuitive, but the theoretical argument is relatively weak. In Eq. (4), the proposed one-step restoration essentially becomes a learned residual update from the degraded latent. The approximation assumes that the degraded latent can be treated as an intermediate flow state and that the current velocity can approximate the average transport velocity. Neither assumption is directly validated. More importantly, the model is trained with a fixed timestep, while timestep variation is presented as a key advantage for restoration-strength control. Eq. (8) relates the update magnitude to the timestep, but does not guarantee monotonic control because the predicted velocity also depends on t. Fig. 6 provides qualitative examples, but I think more direct evidence is needed to support the claimed controllability.

  2. The degradation prototype bank is trained with explicit labels covering 19 degradation types. While the method performs well without degradation prompts at inference, this is different from handling degradation types unseen during training. I think the current experiments mainly demonstrate effective restoration within the degradation distribution covered by CoTIR-Bench. The real-world examples are encouraging, but remain qualitative. A controlled evaluation on held-out degradation types or significantly different corruption distributions would make the claim of universal restoration more convincing.

Questions for authors
  1. How is the assumption that degraded latents correspond to intermediate flow states justified? Also, are the results in Fig. 6 generated from the same checkpoint trained at t = 0.25? Could the authors provide quantitative evidence that changing t consistently controls restoration strength rather than simply changing the output?

  2. Since the degradation prototype bank is trained using labels for 19 degradation types, how does the model behave on degradations outside these predefined categories? Are there any existing held-out degradation results that could demonstrate this generalization?

  3. Appendix A states that the EM updates keep prototype scales fixed when writing parameters back. Could the authors clarify how the Gaussian variances are learned and how they contribute to prototype retrieval? I think this would help explain the gain over conventional prototypes reported in Appendix Table 5.

3 comments

We sincerely thank Reviewer 5x4X for the detailed and constructive feedback. We address each concern directly below.

R-W1 & Q1: Flow motivation and quantitative restoration-strength control.

Our short-horizon formulation builds on a pretrained time-conditioned generative flow model, rather than learning a restoration mapping from scratch. Degraded observations already preserve substantial scene structure, motivating their treatment as informative intermediate states. The new Appendix A (Fig. 10) illustrates this structural correspondence through Gaussian and degradation interpolations. Endpoint supervision then adapts the pretrained flow backbone to predict the remaining correction with one velocity evaluation.

Importantly, our controllability claim is supported by quantitative evidence, not an assumed monotonicity of Eq. (8). Original Fig. 6 (now Fig. 8) uses the same checkpoint trained at (t=0.25). To verify its behavior systematically, we additionally evaluated this fixed checkpoint on all 2,000 CoTIR-Bench test images, holding the degraded latent and prompt fixed while varying only inference (t).

Across (t={0.05,0.15,0.25,0.35,0.45}), mean latent RMS correction increases monotonically as approximately ({0.10,0.29,0.44,0.54,0.62}), while PSNR and SSIM peak and LPIPS reaches its minimum at (t=0.25). This demonstrates predictable average restoration-strength control using a single trained model, while quantifying the fidelity trade-off. The revised Section 3.2 clarifies the formulation, Sections 5.2–5.3 distinguish training from inference timesteps, and Fig. 9 provides the quantitative evidence.

R-W2 & Q2: Generalization beyond the training degradation distribution.

To directly test generalization beyond CoTIR-Bench, we conducted a new evaluation on 200 images spanning eight previously unseen degradation combinations from an independent external benchmark [1]. These combinations were absent from the restoration training data, and all methods were evaluated without additional fine-tuning.

ProtoFlow achieves the best results among all compared methods across all six metrics under both vague and precise prompts. Compared with CoTIR-4B, ProtoFlow improves PSNR by 0.99 dB under vague prompts and 0.91 dB under precise prompts, while reducing vague-prompt LPIPS from 0.4007 to 0.2896. More importantly, compared with the same one-step model without prototype guidance, ProtoFlow improves PSNR by 1.80 dB and 1.41 dB, respectively, with consistent improvements across all six metrics.

The key comparisons are summarized below; the complete comparison with all methods is provided in Appendix E.2 (Table 5).

MethodPSNR ↑SSIM ↑LPIPS ↓CLIP-IQA+ ↑Q-Align ↑LIQE ↑
Input18.510.45210.74600.37942.18921.0677
PromptIR19.800.54170.60830.37172.19871.0747
OneRestore†21.250.64560.48700.39522.40081.2690
InstructIR19.890.58590.50740.37752.61461.2320
InstructIR†20.080.59110.50790.39322.60961.2145
DA-CLIP20.890.61760.46780.42962.85961.6523
HYPIR†20.810.62520.32260.48553.28142.2044
LucidFlux18.270.54790.42460.50893.24682.1552
CoTIR-4B21.680.69410.40070.42232.89221.7054
CoTIR-4B†21.850.69840.38350.43002.97271.7955
ProtoFlow (w/o Proto.)20.870.65320.41230.42183.05762.0476
ProtoFlow† (w/o Proto.)21.350.68750.39760.45363.23582.1563
ProtoFlow22.670.70270.28960.51253.43022.4144
ProtoFlow†22.760.70240.29020.50913.42192.4126

† indicates that a precise degradation prompt is provided. “w/o Proto.” denotes the variant without prototype representation.

These experiments establish zero-shot generalization to previously unseen compositions of degradation types, rather than merely performance on the original benchmark. They also isolate the contribution of prototype guidance under this distribution shift. While this setting differs from entirely held-out degradation categories, it directly evaluates an important form of generalization in real-world restoration, where multiple familiar degradations occur in novel combinations. We have added the full quantitative results and qualitative comparisons in Appendix E.2 (Table 5 and Fig. 14).

[1] Zhu K, et al. An Intelligent Agentic System for Complex Image Restoration Problems. ICLR, 2025.

R-Q3: Gaussian variance learning and prototype retrieval.

Our Gaussian prototypes do not use fixed isotropic variances. We maintain two distinct sets of diagonal variances with complementary roles:

  1. Retrieval variances are prototype- and dimension-specific parameters learned through gradient descent. They directly enter the Gaussian log-likelihood in Eq. (11), allowing prototype matching to account for feature dispersion along individual dimensions.

  2. EM variances are estimated from cached features and retained across successive EM updates, guiding component responsibilities and prototype-center updates. They do not overwrite the learned retrieval variances.

Both sets are initialized identically across prototypes but evolve independently through their respective update mechanisms. Section 4.1 and Appendix B.2 now explicitly document these operations.

We further compare Gaussian prototypes against conventional vector-based prototypes under matched encoder, prototype organization, aggregation, injector, losses, and training settings. Gaussian prototypes improve PSNR from 25.77 to 26.34 dB, SSIM from 0.8540 to 0.8741, and LPIPS from 0.0865 to 0.0836. These controlled results demonstrate the benefit of variance-aware prototype learning and retrieval over conventional prototype representations. The comparison is clarified in Appendix E.3 (Table 6).

AuthorsOpenReview ↗

Request to Review Our Rebuttal and Revised Manuscript

Dear Reviewer 5x4X,

Thank you again for your thoughtful comments on timestep control and generalization beyond CoTIR-Bench. We expanded the explanation of the short-horizon formulation and added a fixed-checkpoint evaluation on all 2,000 CoTIR-Bench test images. Varying only the inference timestep increased the mean latent correction magnitude across the tested range. We also evaluated 200 images from an independent benchmark covering eight degradation combinations absent from training. Without further fine-tuning, prototype guidance improved PSNR by 1.80 dB under vague prompts compared with the same one-step model without prototypes. We have clarified that this result concerns unseen combinations of degradation types, and we now explain how the learned diagonal variances contribute to prototype retrieval.

With about one day remaining in the discussion period, we would be grateful if you could review our rebuttal and revised manuscript and let us know whether the revisions address your concerns. We would appreciate your reconsidering your assessment. If any concerns remain or new questions arise, we would be happy to respond.

Best regards,

The Authors

Reviewer_zaV6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes ProtoFlow for universal image restoration under known or unspecified degradations. It adapts a pretrained FLUX.2-4B model to predict a clean latent from the degraded latent with one velocity evaluation. Two banks of Gaussian prototypes provide degradation and semantic conditioning: the former is trained with multi-label degradation supervision, while the latter aligns paired degraded and clean images through a contrastive objective. A subsequent restoration stage freezes the prototype conditioner and trains an injector and LoRA adapters. Experiments on the 2,000-image CoTIR-Bench test set report better full-reference quality than CoTIR-4B, comparable no-reference scores, and lower inference latency. The paper also presents prototype ablations and qualitative examples of adjusting restoration intensity through the timestep.

Critical strengths
  1. The reported quality–latency trade-off is compelling within the evaluated setting. Without precise degradation prompts, ProtoFlow improves PSNR from 24.52 to 26.34 dB and LPIPS from 0.1206 to 0.0836 relative to CoTIR-4B, while reducing latency from 4.602 to 0.594 seconds for the stated 1024 × 1024 input size (Table 1). The result with precise prompts is also strong. The no-reference scores are competitive rather than uniformly better, which gives a more balanced picture of the contribution. If the inference configurations are comparable, this is a useful practical improvement for a large generative restorer.

  2. The paper provides substantive evidence for prototype conditioning, including a comparison that is easy to miss if one reads only the main table. In Table 2, the one-step model without either bank reaches 25.02 dB under vague prompts; the semantic-only and degradation-only variants reach 25.65 and 25.80 dB, and the combined model reaches 26.34 dB. Appendix C.2 also compares conventional and Gaussian prototypes, reporting 25.77 versus 26.34 dB. These experiments support more than a generic claim that a larger model works better. They suggest that the conditioning design is responsible for an appreciable part of the improvement, although the precise implementation of those controls still needs clarification.

Critical weaknesses
  1. The flow interpretation and the strength-control claim are not adequately connected to the stated training procedure. Section 3.2 treats the current state as approximately the degraded latent, and Section 5.1 gives a default training timestep of 0.25. At a fixed nonzero timestep, minimizing the endpoint loss in Eq. (5) is compatible with learning an ordinary restoration mapping: the velocity can simply encode the required residual divided by that timestep. This does not establish that changing the timestep follows a short transport trajectory or predictably changes restoration strength. Eq. (8) also depends on how the velocity norm changes with time; it does not establish monotonic correction by itself.

  2. Figure 6 does show useful qualitative changes, so I would not dismiss the control behavior outright. However, Table 3 is described as a study of the timestep used to train the model, while the discussion says that larger timesteps move the starting state closer to random noise. That statement is difficult to reconcile with an input state that remains the same degraded latent. I need to know whether Table 3 retrains separate models, whether training samples timesteps or perturbs the latent, and whether Figure 6 changes only the inference timestep of one checkpoint. This distinction matters to the claim that one trained model provides a principled intensity controller. Earlier work already establishes one-step restoration with controllable outputs, so the contribution here needs to be identified more specifically than the existence of a one-step update.

  3. The prototype mechanism is supported empirically but not specified consistently enough to assess the claimed explanation. Eq. (18) replaces the retrieved degradation token with a learnable null embedding for precise prompts. Nevertheless, the degradation-only variant improves the precise-prompt PSNR from 25.35 to 26.11 dB in Table 2. This could have a valid explanation—for example, a changed training configuration or an additional learned embedding—but the paper does not explain what differs when the retrieved degradation token is bypassed. The result therefore cannot yet be interpreted straightforwardly as a benefit from retrieving degradation prototypes in that branch.

  4. There is a related uncertainty about the Gaussian parameterization. Appendix A.2 explicitly says that EM estimates variances but keeps the bank scales fixed when writing parameters back. This does not rule out learning scales through another optimizer, and Table 5 does provide a Gaussian-versus-conventional comparison. What remains unclear is which variance parameters actually change, how they are initialized, and whether the conventional control shares the encoder, losses, training budget, and injector. These details determine whether the observed gain supports modeling intra-prototype variation or a broader change in training and conditioning. Since prototype guidance is the main additional mechanism, I regard this as a material issue rather than a request for a larger ablation grid.

Questions for authors
  1. Please specify the training distribution of timesteps and the exact construction of the latent presented to the restorer. Does Table 3 use separately trained checkpoints? In Figure 6, is one checkpoint evaluated at different timesteps with every other input held fixed? A quantitative correction-magnitude or task-quality curve for that checkpoint would help resolve the strength-control claim.

  2. What changes between the no-prototype, semantic-only, degradation-only, and combined configurations under precise prompts, given the null-token rule in Eq. (18)? Please also clarify which Gaussian scale parameters are optimized outside EM and which settings are held fixed in the conventional-prototype comparison.

  3. Were CoTIR-4B and ProtoFlow timed with the same numerical precision, image/token resolution, batching, attention implementation, and device configuration? Figure 5 reports 4.356 seconds for five CoTIR transformer evaluations and 0.343 seconds for the proposed evaluation, so the improvement is larger than the reduction in evaluation count alone. Identifying the additional configuration or architectural differences would make the reported speedup interpretable.

3 comments

We appreciate the reviewer's recognition of ProtoFlow's substantial improvements in restoration quality, efficiency, and prototype conditioning. We address the technical concerns below with additional experimental evidence and clarifications in the revised manuscript.

R-W1 & Q1 1): Flow formulation and quantitative strength control.

Our formulation leverages a pretrained, time-conditioned flow backbone to learn a short-horizon correction from the degraded latent toward the clean target. As illustrated in the new Appendix A (Fig. 10), degraded observations preserve scene structure similarly to intermediate generative states, motivating this formulation. Although fixed-timestep endpoint training can be expressed as residual prediction, this does not preclude meaningful timestep-dependent restoration behavior. The pretrained backbone retains timestep conditioning, while (t) also explicitly scales the predicted update.

Crucially, we do not rely on Eq. (8) to guarantee monotonic control. We directly evaluated the same checkpoint trained at (t=0.25) on all 2,000 CoTIR-Bench test images, varying only the inference timestep. Across (t={0.05,0.15,0.25,0.35,0.45}), mean latent RMS correction increases monotonically as approximately ({0.10,0.29,0.44,0.54,0.62}), while PSNR and SSIM peak and LPIPS reaches its minimum at (t=0.25). These results establish predictable average restoration-strength control with a single checkpoint over the evaluated range, rather than merely showing visually different outputs. We have clarified the formulation in Section 3.2 and added quantitative evidence in Section 5.3 and Fig. 9.

R-W2 1) & Q1 2): Training timesteps and inference settings.

The restorer always receives the encoded degraded image without additional noise, during both training and inference. Table 3 compares five independently trained models, each using a fixed timestep without timestep sampling. In contrast, Figs. 8–9 use a single checkpoint trained at (t=0.25), with the input latent and prompt fixed while varying only inference (t). Therefore, changing (t) affects the velocity conditioning and update scale, not the starting latent's distance from noise or the clean target. The revised Sections 5.2–5.3 explicitly distinguish these settings and remove the previous ambiguity.

R-W2 2): Distinction from existing controllable one-step restoration.

Our primary contribution is prototype-guided universal restoration under diverse and mixed degradations without requiring explicit degradation descriptions, rather than introducing one-step controllability alone. The complementary Gaussian degradation and semantic banks provide structured guidance for identifying degradations while preserving content.

Unlike OFTSR [1] and TADSR [2], which develop controllability through specialized trajectory-constrained or time-aware distillation, ProtoFlow exploits the existing timestep conditioning of a pretrained flow backbone after fixed-timestep restoration fine-tuning, without an explicit restoration-strength control objective. Controllability is thus an additional capability of our prototype-guided framework, whose primary advantages are restoration accuracy, prompt robustness, and efficiency.

[1] Yuanzhi Zhu et al. OFTSR: One-Step Flow for Image Super-Resolution with Tunable Fidelity-Realism Trade-offs. ICLR, 2026.

[2] Tianyi Zhang et al. Time-Aware One Step Diffusion Network for Real-World Image Super-Resolution. CVPR, 2026.

R-W3 & Q2 1): Precise-prompt ablation and joint training.

Each Table 2 configuration is jointly trained with vague and precise prompts sampled at a 1:1 ratio according to Eq. (18), sharing restoration parameters. Under precise prompts, the no-prototype variant provides no prototype guidance; the semantic-only variant provides semantic guidance; the degradation-only variant uses a learnable null token; and the combined variant uses semantic guidance together with the null token.

During vague-prompt training, degradation retrieval updates the shared restoration parameters subsequently used under precise prompts. Thus, the 25.35→26.11 dB precise-prompt improvement is consistent with benefits from joint training, not direct degradation-prototype retrieval during precise-prompt inference. We have clarified this protocol in Section 4.2.

R-W4 & Q2 2): Gaussian variance learning and ablation fairness.

Our implementation maintains two distinct sets of diagonal variances. Retrieval variances are prototype- and dimension-specific parameters optimized by gradient descent and used for Gaussian likelihood matching. EM variances estimate within-component dispersion from cached features and are retained across EM updates to guide responsibilities and prototype-center updates. Both sets are initialized identically across prototypes, but their subsequent updates are not tied. EM variances do not overwrite retrieval variances. These mechanisms are now explicitly documented in Section 4.1 and Appendix B.2.

For the conventional-prototype comparison, both configurations share the encoder, prototype organization, mixture-weight scheme, aggregation, injector, losses, and training budget. The conventional variant omits Gaussian variance parameters. Under otherwise matched settings, Gaussian prototypes improve PSNR from 25.77 to 26.34 dB and SSIM from 0.8540 to 0.8741, while reducing LPIPS from 0.0865 to 0.0836. These results support the effectiveness of the variance-aware prototype formulation beyond conventional vector-based prototypes. We have clarified the comparison in Appendix E.3 (Table 6).

R-Q3: Runtime fairness and sources of acceleration.

CoTIR-4B and ProtoFlow were timed under matched device, numerical precision, batch size, input resolution, text-processing settings, attention implementation, and timing boundaries.

The transformer-stage acceleration from 4.356 s to 0.343 s (12.7×) combines two measurable factors: 5× fewer velocity evaluations and approximately 2.54× lower runtime per evaluation. The latter reflects the combined architectural differences: replacing CoTIR's CoT adapter with lightweight prototype conditioning and avoiding concatenation of additional image-conditioning tokens with the noisy latent. At 1024×1024 resolution, ProtoFlow reduces visual tokens from 8,192 to 4,096 per transformer evaluation. Therefore, the speedup is explained by both fewer evaluations and lower per-evaluation computational cost, rather than attributing the entire gain to one-step inference. These factors are clarified in Section 5.3 and Fig. 6.

AuthorsOpenReview ↗

Request to Review Our Rebuttal and Revised Manuscript

Dear Reviewer zaV6,

Thank you again for your constructive feedback on the flow formulation and the evidence for timestep control. We clarified that the training ablation compares independently trained fixed-timestep models, whereas the new inference experiment uses a single checkpoint and varies only the timestep, holding the degraded latent and prompt fixed. Across all 2,000 test images, the mean latent correction magnitude increased monotonically over the tested range, with the best full-reference scores at t=0.25t=0.25 (Fig. 9). We also clarified how ProtoFlow’s prototype guidance distinguishes it from prior controllable one-step methods, and expanded the explanations of joint prompt training, Gaussian variance learning, and runtime attribution.

With about one day remaining in the discussion period, we would be grateful if you could review our rebuttal and revised manuscript and let us know whether the revisions address your concerns. We would appreciate your reconsidering your assessment. If any concerns remain or new questions arise, we would be happy to respond.

Best regards,

The Authors

Reviewer_yJm7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ProtoFlow adapts a pretrained FLUX.2-4B model for universal image restoration using a single velocity evaluation from the degraded image latent. It introduces degradation and semantic prototype banks, pretrained using degradation labels and paired low-quality/clean images, respectively. Retrieved prototype representations condition the restorer through an injector, while LoRA adapters are trained with latent reconstruction and pixel-space perceptual and structural losses. On CoTIR-Bench, the method reports better full-reference restoration metrics than CoTIR-4B, a smaller gap between vague and precise prompts, and approximately 7.7 times faster inference. The paper also demonstrates restoration-strength adjustment through the timestep.

Critical strengths
  1. A promising restoration-quality and efficiency trade-off. Table 1 reports an improvement over CoTIR-4B from 24.52 to 26.34 dB PSNR and from 0.1206 to 0.0836 LPIPS under vague prompts, while reducing runtime from 4.602 to 0.594 seconds for 1024 × 1024 inputs. The corresponding precise-prompt results also improve. These gains make the approach practically interesting, even though it is not the fastest method in the table and does not lead every no-reference metric.

  2. Useful evidence that the two prototype branches contribute to restoration. In Table 2, combining both banks improves vague-prompt PSNR from 25.02 to 26.34 dB and LPIPS from 0.0975 to 0.0836. Table 5 also reports better results with Gaussian than with conventional prototypes. The PSNR gap between precise and vague prompts decreases from 0.67 dB for CoTIR-4B to 0.28 dB for ProtoFlow. Together, these results support the usefulness of structured conditioning when explicit degradation descriptions are unavailable.

Critical weaknesses
  1. The implementation of the Gaussian prototypes and their role in the ablations are insufficiently specified. Section 4.1 motivates Gaussian prototypes through their ability to represent intra-prototype dispersion. Appendix A.2, however, states that the current implementation keeps bank scales fixed during EM write-back and updates only the means and priors. This does not establish whether scales are initialized differently, optimized by gradient descent, or held fixed throughout training. If all covariances are the same isotropic matrix, the Gaussian matching score reduces to a scaled squared-distance score plus a constant, so learned dispersion would not explain the improvement. Table 5 is encouraging, but the conventional-prototype control is not described sufficiently to isolate this mechanism. There is a related ambiguity in the degradation-bank ablation: Eq. (18) replaces the retrieved degradation token with a learnable embedding for precise prompts, yet Table 2 reports a sizeable precise-prompt gain from the degradation-only configuration. Joint training could explain this, but the training schedule and parameter sharing need to be stated. These issues limit attribution of the improvements to the proposed prototype design.

  2. The timestep interpretation and the extent of the efficiency claim need better-controlled support. Section 3.2 uses the degraded latent as the current state, but Section 5.3 explains larger training timesteps as placing the starting point closer to random noise. If the input remains the same degraded latent, changing the timestep does not itself change its distance to the clean target or to noise. Equation (8) also does not guarantee increasing correction magnitude because the velocity depends on the timestep. Figure 6 provides qualitative evidence of control, but not a quantitative test across degradations. For runtime, Figure 5 reports a transformer-stage reduction from 4.356 to 0.343 seconds, about 12.7 times, although the number of evaluations decreases fivefold; the paper also changes the conditioning architecture. The measured end-to-end speedup is useful, but the paper should separate its causes and comparison settings more carefully.

Questions for authors
  1. How are the prototype variances initialized and updated? Are they shared, prototype-specific, or dimension-specific? Are they optimized by gradients outside EM? Please specify how the Table 5 control matches the Gaussian version in normalization, score temperature, losses, EM updates, and parameter budget.

  2. What exact latent is supplied to the restorer during training and inference at each timestep? Does Table 3 compare separately trained models? Does Figure 6 use one fixed checkpoint? Please reconcile the statement about approaching random noise with the degraded-latent input, and provide correction-magnitude or degradation-specific fidelity curves for the fixed checkpoint if available.

  3. How are vague and precise prompts sampled during training, and which parameters are shared? In Table 2, what differs between the baseline and degradation-only configuration under precise prompts, given Eq. (18)? This clarification would determine whether the result reflects shared training benefits or direct degradation-prototype conditioning.

  4. Were all runtimes measured with the same device, precision, batch size, attention implementation, text-embedding caching, and timing boundaries? What accounts for the approximately 12.7-fold transformer-stage reduction? A matched one-step/five-step comparison would help distinguish fewer evaluations from conditioning and implementation differences.

2 comments

We sincerely appreciate Reviewer yJm7's valuable feedback. Our responses to the weaknesses and questions are listed below.

R-W1 1) & Q1 1): Gaussian variance initialization and updates.

Our implementation does not rely on fixed isotropic covariances. We maintain two distinct sets of diagonal variances. Retrieval variances are prototype- and dimension-specific parameters optimized through gradient descent, directly affecting Gaussian likelihood matching. EM variances estimate within-component dispersion from cached features and are retained across successive EM updates to guide responsibilities and prototype-center updates. Both sets are initialized identically across prototypes but updated independently thereafter; EM variances never overwrite the learned retrieval variances. These mechanisms are now explicitly described in Section 4.1 and Appendix B.2.

R-W1 2) & Q1 2): Fairness of the conventional-prototype ablation.

The conventional-prototype control uses vector-based representations without Gaussian variance parameters. Both variants retain the same feature encoder, prototype organization, mixture-weight scheme, aggregation, injector, losses, and training budget. The Gaussian configuration achieves 26.34 dB versus 25.77 dB PSNR under vague prompts, supporting the value of variance-aware prototype fitting and retrieval under otherwise matched settings. We have clarified the comparison in Appendix E.3 (Table 6).

R-W1 3) & Q3: Precise-prompt improvements in Table 2.

Each ablation configuration is jointly trained with vague and precise prompts sampled at a 1:1 ratio, sharing the restoration parameters across both settings. As defined in Eq. (18), degradation prototypes are retrieved for vague prompts but replaced with a learned null embedding for precise prompts. Thus, the precise-prompt improvement from 25.35 to 26.11 dB can arise through shared-parameter learning from vague-prompt training, not direct degradation-prototype retrieval during precise-prompt inference. The joint-training protocol is now clarified in Section 4.2.

R-W2 1) & Q2: Timestep interpretation and quantitative strength control.

During both training and inference, the restorer receives the encoded degraded image without added noise. Changing the timestep affects the velocity conditioning and update scale, not the input latent. Table 3 compares separately trained fixed-timestep models, whereas Figs. 8–9 evaluate a single checkpoint trained at t=0.25.

To directly validate controllability, we evaluated this fixed checkpoint on all 2,000 CoTIR-Bench test images, varying only the inference timestep. Across t={0.05, 0.15, 0.25, 0.35, 0.45}, mean latent RMS correction increases approximately as {0.10, 0.29, 0.44, 0.54, 0.62}, while PSNR and SSIM peak and LPIPS reaches its minimum at t=0.25. This establishes empirical restoration-strength control over the tested range, rather than assuming monotonicity from Eq. (8). These results are included in Sections 5.2-5.3 and Fig. 9.

R-W2 2) & Q4: Runtime comparison and acceleration attribution.

The runtime comparisons use matched device, precision, batch size, image resolution, text-processing settings, attention implementation, and timing boundaries.

The transformer-stage acceleration from 4.356 s to 0.343 s (12.7×) combines 5× fewer velocity evaluations with approximately 2.54× lower cost per evaluation. The latter reflects architectural differences: ProtoFlow replaces CoTIR's CoT adapter with lightweight prototype guidance and avoids concatenating additional image-conditioning tokens with the noisy latent. At 1024×1024 resolution, this reduces the transformer visual tokens from 8,192 to 4,096. Therefore, the speedup is attributable to both fewer evaluations and reduced per-evaluation computation, rather than the step reduction alone. The revised Section 5.3 explains these factors, and Fig. 6 presents the runtime breakdown.

AuthorsOpenReview ↗

Request to Review Our Rebuttal and Revised Manuscript

Dear Reviewer yJm7,

Thank you again for your careful review and for asking us to clarify what drives ProtoFlow’s gains. In the revised manuscript, we explain how the retrieval variances are learned separately from the EM variance estimates and document which settings are held fixed in the Gaussian prototype ablation. We also clarify how sharing restoration parameters during joint training can benefit precise-prompt performance, where degradation retrieval is bypassed. To quantify restoration-strength control, we varied only the inference timestep of a single checkpoint across all 2,000 CoTIR-Bench test images. Finally, we revised the runtime breakdown to distinguish the savings from fewer velocity evaluations and the lower cost of each evaluation.

With about one day remaining in the discussion period, we would be grateful if you could review our rebuttal and revised manuscript and let us know whether the revisions address your concerns. We would appreciate your reconsidering your assessment. If any concerns remain or new questions arise, we would be happy to respond.

Best regards,

The Authors

Comments on the paper

1 comment

Dear Reviewers,

We appreciate your recognition of ProtoFlow's restoration quality, efficiency, and prototype design. In the revised manuscript, we clarified the Gaussian variance learning, ablation fairness, joint prompt training, and runtime attribution. We also added three evaluations: 1) a fixed-checkpoint timestep evaluation on all 2,000 test images showed that mean correction strength increased monotonically across the tested timesteps; 2) An external evaluation on eight unseen degradation combinations showed gains over the prototype-free baseline and other SOTA methods; 3) A user study with 50 participants showed higher mean ratings than others for visual quality, fidelity, and absence of hallucinations.

As the discussion period draws to a close, we would be grateful if you could reconsider your ratings in light of the revised manuscript and the new results. We welcome any further questions or comments.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.