{"id":"4a2de4b3-cb3b-4457-b32b-359a2084ed81","arxiv_id":"2501.13134","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"UniRestore adapts a Stable Diffusion restoration model to simultaneously improve perceptual quality, classification accuracy, and segmentation mIoU by restoring encoder features and using task-specific prompts.","lead":"UniRestore is a single image restoration model built on Stable Diffusion that improves both human-perceived quality and downstream task performance. It adds two modules, one that repairs features inside the autoencoder and one that adapts those features to different tasks through lightweight prompts.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Extendability claim is under-supported: Table 6 lacks a no-new-prompt control and reports no original-task metrics after adding detection; 'any task' (Fig. 1c) is extrapolation.","rationale":"After reading the paper in good faith, the core PIR/TIR comparisons are the strongest part of the work: Table 1 shows consistent PSNR/SSIM gains over PromptIR and DiffUIR across seen and unseen datasets, and Tables 2-3 show large margins in classification and segmentation, with ablations (Tables 4-5) supporting the contribution of both CFRM and TFA. The most fragile link in the central argument is not the per-task performance but the extensibility claim, which is explicitly advertised in Figure 1c and Section 3.4. The reader's weakest_assumption already targets this; the present stress test sharpens it by identifying a missing control (an existing prompt on the new task) and missing interference measurements. A single object-detection result on RTTS, without these controls, cannot establish prompt-only adaptation to 'any task.' This concern does not overturn the paper; it makes the acceptance conditional on a stronger extensibility evaluation. The DIP average inconsistency in Table 1 (e.g., the DIP row PSNR values 18.47, 22.65, 21.30, 19.03, 15.41, 23.08 average to 19.99, not the reported 17.13) is a separate correctness flag that further supports requiring a reproducibility pass, but it is not the main load-bearing issue.","tokens_in":14628,"tokens_out":10479,"duration_ms":104118,"concrete_test":"Run the Table 6 detection setup on RTTS with three UniRestore conditions, all with CFRM, Controller, SC-Tuner, and TFA frozen: (a) a freshly initialized prompt trained only on detection data, (b) the existing PIR prompt (or classification prompt) with no training on detection data, and (c) no prompt / default latent-only decoding. If (a) is not clearly better than (b), the improvement is not due to task-specific prompt adaptation. In the same run, evaluate DIV2K PSNR/SSIM, ImageNet classification ACC, and Cityscapes mIoU under the detection prompt; if any drop exceeds run-to-run variability (report at least 3 seeds with error bars), the 'unified' claim is weakened. Optionally repeat with a second task requiring different spatial granularity (e.g., depth estimation or keypoint detection) to test the 'any task' extrapolation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.4 claims that adding a TIR task requires only a new prompt vector, with CFRM, Controller, SC-Tuner, and the shared TFA frozen. The only evidence is Table 6: on RTTS object detection, fine-tuning a new prompt yields 58.06 mAP vs 45.63 for LQ, 54.29 for DIP*, and 50.61 for PromptIR*. This does not establish the mechanism. There is no control using UniRestore with an existing prompt (e.g., the PIR prompt) on RTTS; without it, the gain could come from the frozen restoration backbone and TFA rather than from the new task-specific prompt. Moreover, after adding the detection prompt, the paper does not re-report PIR (DIV2K PSNR/SSIM), classification (ImageNet ACC), or segmentation (Cityscapes mIoU), so it is unknown whether the 'unified' model retains original-task performance. Since the 'any task' generalization in Figure 1c is a stated contribution and the basis for the paper's adaptability and efficiency claims, this is a load-bearing gap: a single successful prompt on one detection benchmark, without controls, cannot support prompt-only extensibility to arbitrary downstream tasks.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents UniRestore, a unified image restoration model built on a latent diffusion prior (Stable Diffusion) with two new components: a Complementary Feature Restoration Module (CFRM) that reconstructs encoder features from degraded inputs, and a Task Feature Adapter (TFA) that fuses these restored features into the decoder under the control of lightweight task prompts. Training proceeds in two stages: first, the CFRM, Controller, and SC-Tuner are trained with a PIR feature-matching loss and latent alignment loss; second, the TFA and task prompts are trained with a multi-task loss combining MSE for PIR, cross-entropy for ImageNet classification, and cross-entropy for Cityscapes segmentation, while earlier modules are frozen. The experiments compare UniRestore against DIP, URIE, NAFNet, PromptIR, DiffBIR, and DiffUIR on PIR benchmarks (DIV2K, Rain100L, RESIDE, UHDSnow, Noise, GoPro) and TIR benchmarks (ImageNet/CUB classification, Cityscapes/FoggyCityscapes/ACDC segmentation), plus an object-detection extension on RTTS. UniRestore reports the best PSNR/SSIM and highest ACC/mIoU on most metrics, and the ablations isolate contributions of CFRM and TFA.","tokens_in":14946,"tokens_out":8126,"duration_ms":72354,"significance":"If the central claims hold, UniRestore is a strong engineering contribution: it demonstrates a single diffusion-based restoration model that serves both perceptual and task-oriented objectives, with a parameter-efficient prompt-based mechanism for adding new tasks. The experimental design includes unseen-domain TIR benchmarks and a multi-task fine-tuning baseline protocol, which is valuable. The main caveats are that the headline extensibility claim rests on a single prompt-only experiment without the controls necessary to attribute the gain to the task prompt, and the primary PIR table contains arithmetic errors in its average columns. Correcting these issues would make the paper a solid contribution to the restoration and low-level vision community.","major_comments":[{"comment":"The reported Average PSNR and SSIM values are arithmetically inconsistent with the six per-dataset entries. For DIP, the listed PSNR values (18.47, 22.65, 21.30, 19.03, 15.41, 23.08) average to 19.99, not 17.13, and the SSIM values average to 0.669, not 0.5734. The DIP* row has the same problem (PSNR average 18.99 vs. reported 16.28). Since the average column is used to support the claim that UniRestore achieves the best overall performance, these errors must be corrected and the table regenerated before the comparison can be trusted.","section":"Table 1"},{"comment":"The claim that adding a new TIR task requires only a new task-specific prompt is load-bearing for the paper's extensibility and efficiency contributions, but the evidence is insufficient. Table 6 reports a single object-detection result on RTTS (58.06 mAP) against DIP and PromptIR, yet it provides no control using UniRestore with an already-learned prompt (e.g., the PIR prompt) on RTTS, and it does not re-report the original-task metrics (DIV2K PSNR/SSIM, ImageNet ACC, Cityscapes mIoU) after the detection prompt is added. Without these controls, the gain cannot be attributed to the new task prompt as opposed to the frozen restoration backbone and TFA, and we cannot know whether prior-task performance is retained. In addition, one detection benchmark is too narrow to support the 'Any Task' generalization in Figure 1(c).","section":"Section 3.4, Table 6"}],"minor_comments":[{"comment":"In the Abstract, 'expertments' should be 'experiments'; in the same paragraph, 'adapability' should be 'adaptability'.","section":"Abstract"},{"comment":"Equation (2) defines L_CFRM as a sum of vector differences; a norm (e.g., L1 or L2) is needed to produce a scalar loss, and the text should specify which norm is used.","section":"Equation (2)"},{"comment":"In Table 3, the header 'Cityscapes [47]' is wrong: the Cityscapes dataset should be cited as [11], while [47] refers to FoggyCityscapes; the current citation is misleading.","section":"Table 3"},{"comment":"In Section 5.3, the phrase 'enhances classification accuracy by 20.01%' should be 'by 20.01 percentage points,' since the metric is absolute accuracy.","section":"Section 5.3"},{"comment":"The term 'seen dataset' should be clarified: the model has seen the same image distribution during training but not the specific test images; this is standard but deserves a sentence to avoid confusion.","section":"Sections 4.2 and 5.2"},{"comment":"No code or project link with code is provided, which limits reproducibility; the project page is mentioned but one would expect a code release.","section":"Reproducibility"},{"comment":"In Table 5, UniRestore's PIR PSNR (24.32) is lower than Multi-TFA (25.48), so the text's statement that UniRestore delivers 'comparable' PIR performance should be refined or explained.","section":"Table 5"},{"comment":"Figure 2 is very dense; adding explicit labels for Stage 1 versus Stage 2 freezing would make the training pipeline easier to follow.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely acceptable if the authors provide corrected Table 1 averages and a proper extendability experiment with the missing controls, including re-reporting of original-task metrics. The absence of code and the table arithmetic error suggest a need for a careful revision cycle, but the central method and most experimental results are credible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I think this paper is worth serious referee time. The core contribution is a clean combination: CFRM restores degraded VAE encoder features, TFA fuses those restored features into the decoder under a per-task prompt vector, all on top of a frozen Stable Diffusion backbone. That specific arrangement is new relative to the cited URIE/DIP/DiffUIR work, and the ablations in Tables 4 and 5 give real evidence that both modules matter. The PIR numbers are consistently ahead on the seen set and most unseen sets, and the TIR gains on classification and segmentation are large (71.65 acc vs 72.80 HQ on ImageNet; 66.05 mIoU vs 75.64 HQ on Cityscapes). For a multi-task restoration system covering three tasks, the results look credible.\n\nThe soft spots are real but mostly fixable. First, the DIP row in Table 1 has an arithmetic problem: the listed PSNR values average to about 20.0, not 17.13. That error is small in isolation but it makes me want to check every number, and the paper gives no error bars or significance tests, so some of the close margins (e.g., DiffUIR on Noise) could be noise. Second, there is no code or supplementary material, which matters for a system built on stable diffusion plus three custom modules. Third—the big one—the 'any task' extensibility claim in Section 3.4 is load-bearing and under-supported. The evidence is Table 6: one detection benchmark (RTTS) with a new prompt, no control where an existing prompt (say, the PIR prompt) is used on the detection task, and no re-reporting of the original three tasks after adding detection. Without those controls you cannot tell whether the gain comes from the new task-specific prompt or just from the frozen backbone+TFA. That is not fatal to the paper's main claim—the unified three-task model is the actual contribution—but the paper sells 'any task' as a headline, and that part needs another experiment or a scaled-back claim.\n\nI also agree with the reader that the citation pattern is fine; the related work covers the main baselines. The method is coherent and the writing is clear.\n\nWho should read this: anyone building restoration models for downstream vision, especially people combining diffusion priors with task adapters. It deserves peer review. My recommendation: accept with major revisions—require code release, fix the table arithmetic, add significance or error bars, and either add the missing controls for the detection-prompt experiment or explicitly restate the extensibility claim as a hypothesis.","headline":"Solid unified restoration system with a genuinely new adapter combination, but the 'any task' extensibility claim rests on one prompt-tuning experiment without the needed controls.","tokens_in":15443,"tokens_out":2236,"would_cite":true,"duration_ms":21268,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"UniRestore claims that a single diffusion-based model can restore images in a way that is both perceptually appealing to humans and useful for downstream computer-vision tasks, by injecting two lightweight feature modules into a frozen…","keywords":["image restoration","diffusion prior","task-oriented image restoration","perceptual image restoration","feature adaptation","prompt tuning","semantic segmentation","image classification"],"falsifier":"Evaluate the prompt-only extension on a spatially dense, geometry-oriented task such as monocular depth estimation under the same protocol; if the frozen TFA cannot yield competitive results without extra trainable modules, then the paper's 'any task' extensibility claim would be refuted.","tokens_in":14473,"feed_emoji":"🖼️","tokens_out":6494,"duration_ms":54192,"temperature":0.7,"pith_summary":"This paper introduces UniRestore, a unified image restoration model that claims to produce images that are simultaneously good for human viewing and good for downstream computer-vision tasks such as classification and segmentation. The authors argue that perceptual restoration and task-oriented restoration are usually in tension, and that existing methods improve one at the expense of the other. UniRestore avoids the trade-off by starting from a frozen diffusion prior that is already strong at perceptual quality and adding two trainable modules that steer the prior toward task-specific requirements. The paper reports that UniRestore beats all compared PIR and TIR baselines on the seen dataset, reaches classification accuracy within about one point of the high-quality ground truth on ImageNet, and requires only a new prompt vector to be optimized when a new task is added.","feed_headline":"One diffusion model unifies perceptual and task-oriented restoration","feed_subtitle":"Two small feature modules steer a frozen diffusion decoder to serve classification and segmentation without retraining.","key_machinery":"The load-bearing components are the Complementary Feature Restoration Module (CFRM) and the Task Feature Adapter (TFA), both inserted into the frozen latent-diffusion autoencoder. CFRM, placed at each encoder layer, reconstructs clear features from degraded ones through four steps: feature enhancement, intra-group channel attention, inter-group channel integration, and feature recovery. TFA, placed at each decoder layer, uses an LSTM-style prompt-update mechanism: a task-specific learnable vector is updated layer by layer and, together with the CFRM-restored features, is fused into the diffusion features via a tuner operator. This two-module design lets the model keep the diffusion prior fixed and adapt only the lightweight prompt, giving task-specific control while preserving the perceptual quality the prior already provides.","core_discovery":"The central discovery is that the encoder features of the diffusion autoencoder, after being restored by the Complementary Feature Restoration Module (CFRM), carry complementary information that a Task Feature Adapter (TFA) can fuse into the decoder's diffusion features under the control of a per-task prompt vector. The paper shows that with this design, the same frozen diffusion decoder can be adapted to produce restorations that serve both perceptual metrics (PSNR/SSIM) and the accuracy of a recognition network. In experiments, UniRestore achieves 71.65% classification accuracy on degraded ImageNet compared to 72.80% for high-quality inputs, and 66.05 mIoU on degraded Cityscapes compared to 75.64 for high-quality inputs, with larger gains on unseen datasets. The paper also demonstrates that adding object detection as a third downstream task requires only optimizing a new prompt, supporting the claim of extensibility.","pith_inferences":["The paper's 'any task' claim is extrapolated from a single additional detection experiment; a more stringent test would be a task with radically different feature demands, such as monocular depth estimation or instance segmentation, using the same frozen TFA and only a new prompt.","The prompt vector likely acts as a soft selector for how much of the CFRM-restored features to blend in at each decoder layer, which suggests the approach could be interpreted as a form of input-conditional feature merging; comparing against other parameter-efficient adaptation methods would clarify the source of the gain.","Because the diffusion prior is frozen, the model's ceiling for tasks that require hallucinating fine structure (e.g., text or faces) may be limited by the prior; a testable extension would be to unfreeze the decoder for a new task and measure the resulting PIR/TIR trade-off."],"forward_implications":["If the claims hold, one unified restoration model can replace separate PIR and TIR pipelines, reducing system redundancy in applications such as autonomous driving that need both human-facing output and machine-facing input.","Adding a new downstream task reduces to optimizing a new prompt vector on that task's data, so extending UniRestore to new tasks is cheaper than retraining a full model.","The success of injecting CFRM-restored encoder features suggests that degradation compensation at the encoder level is an effective place to reconcile perceptual and task objectives.","UniRestore's gains on unseen datasets (e.g., 20.01% accuracy improvement on CUB for ResNet-50) indicate that the approach generalizes beyond the synthetic degradations seen in training.","The shared TFA across tasks implies that multi-task training can transfer knowledge between tasks, as the paper notes when comparing to task-specific TFA variants."],"supporting_citations":[{"why":"Provides the frozen Stable Diffusion prior and autoencoder that UniRestore augments; the backbone whose perceptual quality is the starting point.","marker":"[45]"},{"why":"The Controller architecture that injects control signals into the denoising U-Net; reused and trained in Stage 1 to align latent features.","marker":"[74]"},{"why":"Supplies the SC-Tuner and tuner operator used both in the controller and inside TFA for feature transformation.","marker":"[23]"},{"why":"URIE, the universal image enhancement baseline for visual recognition; UniRestore's task-oriented approach is compared against it and extends its objective.","marker":"[52]"},{"why":"The degradation synthesis procedure used to create the 15-type degraded training data for both PIR and TIR task experiments.","marker":"[16]"},{"why":"DiffUIR, a diffusion-based universal restoration baseline that UniRestore compares against on perceptual metrics and downstream tasks.","marker":"[77]"},{"why":"PromptIR, a prompt-based all-in-one restoration baseline representing the PIR approach that UniRestore improves upon in task-oriented settings.","marker":"[42]"}],"fun_headline_variants":["UniRestore: one diffusion prior for both human and machine vision","Bridging perceptual and task-oriented restoration with a diffusion prior","Restore images for humans and AI: UniRestore's dual purpose","Adaptive diffusion features balance quality and task accuracy","UniRestore: new tasks via new prompts on a frozen diffusion decoder"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that a single frozen TFA and the frozen diffusion decoder can be steered to any new downstream task using only a freshly optimized prompt vector, yet this is validated on exactly one additional task beyond classification and segmentation.","fun_headline_variants_meta":{"raw":{"variants":["UniRestore: one diffusion prior for both human and machine vision","Bridging perceptual and task-oriented restoration with a diffusion prior","Restore images for humans and AI: UniRestore's dual purpose","Adaptive diffusion features balance quality and task accuracy","UniRestore: new tasks via new prompts on a frozen diffusion decoder"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001078,"raw_usage":{"total_tokens":4515,"prompt_tokens":951,"completion_tokens":3564,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":3474}},"tokens_in":567,"tokens_out":3564,"duration_ms":27531,"temperature":1.0,"reasoning_tokens":3474,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:53:24.307788+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the prompt-only extension on a spatially dense, geometry-oriented task such as monocular depth estimation under the same protocol; if the frozen TFA cannot yield competitive results without extra trainable modules, then the paper's 'any task' extensibility claim would be refuted.","supporting_citations":[{"cited_title":"Adding conditional control to text-to-image diffusion models","cited_arxiv_id":null,"evidence_quote":"The Controller architecture that injects control signals into the denoising U-Net; reused and trained in Stage 1 to align latent features."},{"cited_title":"Urie: Universal image enhancement for visual recognition in the wild","cited_arxiv_id":null,"evidence_quote":"URIE, the universal image enhancement baseline for visual recognition; UniRestore's task-oriented approach is compared against it and extends its objective."},{"cited_title":"Selective hourglass mapping for universal image restoration based on diffusion model","cited_arxiv_id":null,"evidence_quote":"DiffUIR, a diffusion-based universal restoration baseline that UniRestore compares against on perceptual metrics and downstream tasks."},{"cited_title":"Promptir: Prompting for all-in- one image restoration","cited_arxiv_id":null,"evidence_quote":"PromptIR, a prompt-based all-in-one restoration baseline representing the PIR approach that UniRestore improves upon in task-oriented settings."}],"review_version":1}