{"id":"a6463032-cd87-4fb9-9db7-822bceb4959c","arxiv_id":"2506.05450","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"A photo-enhancement pipeline uses a vision-language model to classify degradation type and then routes the image to a specialized restoration model, but it only offers anecdotal visual results as evidence.","lead":"This paper describes an automated photo-enhancement pipeline that uses a vision-language model to classify a photo's problem (low resolution, reflections, motion blur, or none) and then routes it to a specialist restoration model. The work matters as a practical system integration, but the paper provides no quantitative evidence for its accuracy claims.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Zero-shot VLM classification accuracy is untested; a misclassification routes images to the wrong restoration model, yet no classification or end-to-end metric is reported.","rationale":"I agree with the reader's verdict and with their identification of the weakest premise: the zero-shot VLM classifier is never evaluated, despite being the routing mechanism on which the entire pipeline depends. The paper's experimental section contains only three illustrative examples, so the claims of 'high accuracy' and 'significant improvements' are claims-without-derivation. I credit the authors for clearly specifying the prompt, the component models, and the routing logic, which makes the missing evaluation concrete and testable; the deployment note in Section V is suggestive but provides no measurable evidence. No machine-checked proof, code, or released dataset is supplied, so there is no independent support to offset the missing experiments. I would keep the verdict at REJECT for the submitted version rather than moving to CONDITIONAL, because the absence of any quantitative evaluation is not a minor gap but the entirety of the evidence for the central claim. A revised submission with the proposed benchmark, a confusion matrix, and end-to-end quality metrics would justify a different outcome.","tokens_in":7228,"tokens_out":1456,"duration_ms":16614,"concrete_test":"Build a labeled benchmark (e.g., 100 images per class from RealSR or Real-ESRGAN for A, SIR2 or RRW for B, GoPro or RealBlur for C, and clean high-quality images for D). Run the exact Section III.A prompt through Qwen2.5-VL and compute per-class accuracy and the confusion matrix. Then run the full pipeline on correctly vs. incorrectly classified subsets and report the mean change in PSNR/LPIPS relative to the original input. If classification accuracy is below roughly 85% on this benchmark, or if misclassified images have worse mean quality than doing nothing, the routing claim fails.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim (Abstract, Section IV, Conclusion) is that the framework achieves high accuracy in degradation classification and significant visual improvement. The load-bearing premise is that Qwen2.5-VL, used zero-shot with the prompt in Section III.A, reliably categorizes real-world degraded images into A/B/C/D. Section IV reports only three anecdotal image examples, with no classification accuracy, no confusion matrix, and no end-to-end quality metric (PSNR, LPIPS, FID, or user study). Because a wrong label sends the image to the wrong specialist model, the pipeline can actively degrade an image that was acceptable or mildly degraded. The class definitions themselves are fragile: category A is a broad union of noise, blur, and JPEG compression, while category C is a separate motion blur class, and the VLM must also distinguish 'no visible degradation' from subtle real-world artifacts. No ablation isolates the classification error rate or measures the downstream cost of misrouting. The deployment claim in Section V is not accompanied by on-device accuracy or quality data, so it cannot substitute for the missing evaluation. The central assertion is therefore unsupported at the level of internal consistency: the paper states 'Experimental results demonstrate...' without reporting the measurements that would substantiate either half of the claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an automated image enhancement pipeline in which a Vision-Language Model (Qwen2.5-VL) classifies an input image into one of four degradation categories — super-resolution degradation (noise, blur, JPEG), reflection artifacts, motion blur, or no visible degradation — and then routes the image to specialized restoration models: InvSR with Real-ESRGAN for text regions (category A), YOLO/YOSO with LaMa and NAFNet for reflections (category B), and NAFNet with CodeFormer for motion blur (category C). The authors state in the abstract, introduction, and conclusion that experiments demonstrate high classification accuracy and significant visual improvement. Section IV, however, contains only three anecdotal example images and no quantitative metrics of any kind.","tokens_in":7535,"tokens_out":2201,"duration_ms":25950,"significance":"If validated, the framework could be a practically useful industrial system, combining a zero-shot VLM classifier with specialist restorers to avoid manual selection of restoration algorithms. The paper has the merit of assembling existing pretrained components into a concrete, deployable pipeline, and the authors state that the modules have been deployed in OPPO smartphones. The central claim, however, is entirely unquantified: the paper reports no classification accuracy, no confusion matrix, no end-to-end image quality metrics (PSNR, SSIM, LPIPS, FID, or user study), and no comparison against baselines or against simply applying each restoration model unconditionally. Because misclassification actively reroutes an image to the wrong restoration model, the untested VLM classification accuracy is load-bearing, and the paper's experimental section does not support the claimed effectiveness.","major_comments":[{"comment":"The Experiments section contains no quantitative evaluation whatsoever. It reports only three representative images per degradation type in Fig. 5, with no classification accuracy, no confusion matrix, no PSNR/SSIM/LPIPS/FID, and no user study. The sentences 'Experimental results demonstrate...' and 'The results show that our method is effective...' are therefore unsupported. This is load-bearing because the abstract and conclusion both assert high classification accuracy and significant visual improvement, and the paper's contribution is exactly the routing decision made by the VLM.","section":"Section IV"},{"comment":"The zero-shot VLM classification is the central premise of the pipeline, yet its reliability is untested. The prompt categories are ambiguous and overlapping: category A includes 'blur' generally while category C is 'motion blur'; category D requires distinguishing 'no visible degradation' from subtle real-world artifacts. The paper gives no accuracy measurement, no failure analysis, and no discussion of the downstream cost of misrouting. A wrong label sends the image to the wrong specialist model, potentially degrading an image that was already acceptable, so the absence of any classification evaluation leaves the mechanism unvalidated.","section":"Section III.A"},{"comment":"Even if classification were perfect, the paper does not show that chaining the specialist models preserves or improves quality. InvSR, Real-ESRGAN, LaMa, NAFNet, and CodeFormer are each pretrained for their own tasks; combining them (e.g., text-region fusion in Section III.B, inpainting followed by NAFNet in Section III.C, and deblurring followed by face restoration in Section III.D) can introduce artifacts or alter color and texture. No end-to-end quality metric or comparison to the unmodified specialist models is reported, so the claimed 'significant improvements in visual quality' are not demonstrated.","section":"Section IV, Fig. 5"},{"comment":"The deployment claim — 'these three AI-powered enhancement modules have already been deployed in OPPO AI smartphones, where they process tens of thousands of user images daily' — is accompanied only by a YouTube link and no on-device accuracy or quality data. This statement cannot substitute for the missing experimental evaluation, and it is not verifiable from the manuscript. If deployment data exist, they should be reported as measurements; otherwise, the claim should be removed or clearly labeled as an industrial deployment note rather than evidence.","section":"Section V"}],"minor_comments":[{"comment":"The category definitions should be clarified to avoid overlap, particularly the boundary between category A ('blur') and category C ('motion blur'), and between category A ('JPEG compression') and category D ('no visible degradation'). A precise decision rule or examples of borderline cases would help reproducibility.","section":"Section III.A"},{"comment":"The paper calls Qwen2.5-VL 'the leading open-source VLM' without citation or benchmark comparison. This is an unsupported superlative; it should be replaced by a specific reference or removed.","section":"Section II.A and III.A"},{"comment":"The experimental section would benefit from stating the model versions, hyperparameters, and any prompt parsing logic used to convert free-form VLM responses (e.g., 'A', 'Category B', or a sentence) into a routing decision. This information is essential for reproducibility.","section":"Section IV"},{"comment":"Several references are incomplete or inconsistently formatted: reference [4] lacks page numbers, reference [24] has a missing paper title and inconsistent author formatting, and the footnote in Section V contains a raw YouTube URL that should be replaced with a proper citation.","section":"References"}],"recommendation":"reject","confidential_remarks":"The manuscript is an industrial system description with no quantitative evaluation. The central claim of high classification accuracy and significant visual improvement is unsupported by the reported evidence, and the missing evaluation is not a minor presentation issue but the core of the paper's contribution. Adding a proper evaluation with classification accuracy, end-to-end quality metrics, and misrouting analysis would constitute a substantially new manuscript rather than a local revision. I also note that the deployment claim in Section V is not backed by any data in the paper, and the large number of self-citations in the reflection-removal related work does not affect the technical validity but may deserve editorial attention."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a system paper with a nice hook and almost no experimental content. The hook: use a zero-shot VLM (Qwen2.5-VL) to label an image into one of four degradation buckets, then route it to a specialist restoration model. That specific routing idea is new as far as I know, and it is genuinely practical for a phone camera pipeline. The writing is clear, the choice of off-the-shelf models is sensible, and the authors know the literature. The deployment claim (tens of thousands of images a day on OPPO phones) is at least a sign that the system works in practice.\n\nThe soft spot is not a soft spot; it is the load-bearing floor. Section IV is three example images and a paragraph saying 'the results show.' There is no classification accuracy, no confusion matrix, no PSNR/LPIPS, no comparison with running a single generic restoration model, no user study. The abstract and introduction claim 'high accuracy' and 'significant improvements' without reporting either measurement. That is a claim without derivation.\n\nAlso worth noting: the four categories are not clean. Category A is a union of noise, blur, and JPEG compression; category C is motion blur. A blurry image could legitimately go to either A or C, and the VLM is asked to make that call with no calibration data. There is no discussion of what happens when the VLM is wrong. For a degraded photo, misrouting can actively make it worse. The absence of any error analysis means the central premise is untested.\n\nI agree with the reader's take. This is not a bad idea and not a bad engineering description; it is an unfinished paper. If the authors came back with a proper evaluation — classification accuracy on a few hundred labeled examples, end-to-end quality metrics, and an ablation on misrouting cost — I would look at it again seriously. As it stands, the experiments section is a placeholder.\n\nWho is this for? Someone surveying VLM applications for image restoration, or someone building a phone enhancement pipeline. A serious referee would not be able to verify any claim from the current text, so I would not send it out in its current form. But I would not burn the paper; I would invite a resubmission with real numbers.\n\nRecommendation: reject, but make clear that the bottleneck is the missing evaluation, not the approach.","headline":"A practical VLM-routing idea with no experiment to back it; the central accuracy claim is unsupported by the three example images in Section IV.","tokens_in":7996,"tokens_out":2373,"would_cite":false,"duration_ms":26651,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims a zero-shot vision-language model can classify image degradation types and route each photo to a specialist restoration model, producing visibly improved results without retraining.","keywords":["image restoration","vision-language model","zero-shot classification","degradation classification","super-resolution","reflection removal","motion deblurring"],"falsifier":"Take a labeled corpus of degraded images covering the four categories, run the paper's exact prompt through Qwen2.5-VL, and compute a confusion matrix; if classification accuracy is close to chance, or if images routed by the classifier show no improvement over the unprocessed input on a perceptual or reference-based metric, the central claim fails.","tokens_in":7004,"feed_emoji":"📷","tokens_out":7568,"duration_ms":66179,"temperature":0.7,"pith_summary":"The paper proposes a fully automated image-enhancement pipeline in which a vision-language model classifies every input photo into one of four categories: super-resolution degradation, reflection artifacts, motion blur, or no visible degradation. Based on that label, the image is sent to a restoration model built for the corresponding problem, such as a diffusion-based super-resolver, an inpainting model for reflections, or a deblurring network. The authors argue that this design replaces manual selection of restoration algorithms and scales to real-world use because the classifier needs no fine-tuning. The stated experimental conclusion is that the approach achieves accurate degradation classification and clear visual improvement across all three degradation types.","feed_headline":"A vision-language model routes photos to the right restoration tool","feed_subtitle":"Zero-shot classification picks the right fix for each degraded photo, automating photo repair without manual tuning.","key_machinery":"The load-bearing component is the zero-shot classification step, in which Qwen2.5-VL receives a fixed textual prompt and outputs one of four letters, A through D, as a routing decision. That decision determines which specialist model handles the image: InvSR with Real-ESRGAN text fusion for super-resolution, a YOLO/YOSO mask plus LaMa inpainting and NAFNet refinement for reflections, and NAFNet deblurring with CodeFormer face restoration for motion blur. The VLM's classification is what turns a collection of independent restoration models into an automated pipeline.","core_discovery":"The central claim, stated on the paper's own terms, is that Qwen2.5-VL can act as a zero-shot degradation classifier that reliably separates images into four categories, and that routing each category to a dedicated restoration model yields higher-quality outputs than a one-size-fits-all approach. For super-resolution degradation the pipeline combines InvSR with PaddleOCR-detected text regions restored by Real-ESRGAN; for reflection artifacts it detects strong reflections with YOLO and YOSO, inpaints them with LaMa, and refines weak reflections with NAFNet; for motion blur it deblurs with NAFNet and enhances faces with CodeFormer. The paper presents visual examples as evidence and concludes that the framework is scalable, automated, and effective for real-world image enhancement.","pith_inferences":["Editorial inference: the four-way taxonomy puts noise, blur, and JPEG compression under a single label, so an image with only mild compression will be processed by a heavy super-resolution model even though a lighter correction might suffice.","Editorial inference: because the VLM is asked to return a single letter, it has no built-in way to signal uncertainty or mixed degradations; a confidence score or a multi-label output could let the pipeline skip restoration or call two specialists in sequence.","Editorial inference: a quantitative evaluation of the framework could compare end-to-end outputs against the unprocessed input on perceptual metrics or user studies; the paper's visual examples leave the magnitude of the improvement unspecified."],"forward_implications":["If the VLM classifies accurately, users no longer need to diagnose an image's defect or choose a restoration algorithm manually, because the same prompt-driven model makes that choice.","The framework is plug-and-play in the sense that the classifier is used without fine-tuning, so new degradation categories could in principle be added by editing the prompt and attaching another specialist model.","The modular design means each restoration stage can be upgraded independently, so improvements in super-resolution, reflection removal, or deblurring should transfer directly into the pipeline.","The authors state that the three enhancement modules are already deployed in consumer smartphones and process tens of thousands of user images daily, implying the approach is practical at scale."],"supporting_citations":[{"why":"Supplies the zero-shot vision-language classifier (Qwen2.5-VL) that the entire routing decision depends on.","marker":"[13]"},{"why":"Provides InvSR, the diffusion-inversion super-resolution model used for category A restoration.","marker":"[1]"},{"why":"Provides Real-ESRGAN, used to restore text regions that the super-resolution model handles poorly.","marker":"[3]"},{"why":"Provides LaMa, the inpainting model that fills masked strong-reflection regions in category B.","marker":"[6]"},{"why":"Provides NAFNet, used both to refine weak reflections and to perform general motion deblurring.","marker":"[7]"},{"why":"Provides CodeFormer, the face-restoration model applied after deblurring for category C.","marker":"[8]"},{"why":"Provides YOLO, the detector used to locate strong reflection regions before mask generation.","marker":"[4]"},{"why":"Provides YOSO, the segmentation model used to refine the reflection mask at pixel level.","marker":"[5]"},{"why":"Detects text regions in super-resolution images so Real-ESRGAN can restore them separately.","marker":"[2]"}],"fun_headline_variants":["Vision-language model auto-classifies image degradation for targeted fixes","VLM routes each degraded photo to its best restoration model","Zero-shot degradation classifier picks the right repair tool automatically","Automated photo repair: VLM classifies, then specialized models restore","From blur to reflection: VLM chooses the correct restoration pipeline"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire pipeline rests on the assumption that the Qwen2.5-VL vision-language model, used without any task-specific training, will put real-world photos into the correct one of four degradation categories from the prompt alone.","fun_headline_variants_meta":{"raw":{"variants":["Vision-language model auto-classifies image degradation for targeted fixes","VLM routes each degraded photo to its best restoration model","Zero-shot degradation classifier picks the right repair tool automatically","Automated photo repair: VLM classifies, then specialized models restore","From blur to reflection: VLM chooses the correct restoration pipeline"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000377,"raw_usage":{"total_tokens":1967,"prompt_tokens":862,"completion_tokens":1105,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":478,"completion_tokens_details":{"reasoning_tokens":1021}},"tokens_in":478,"tokens_out":1105,"duration_ms":10559,"temperature":1.0,"reasoning_tokens":1021,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:20:27.221522+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a labeled corpus of degraded images covering the four categories, run the paper's exact prompt through Qwen2.5-VL, and compute a confusion matrix; if classification accuracy is close to chance, or if images routed by the classifier show no improvement over the unprocessed input on a perceptual or reference-based metric, the central claim fails.","supporting_citations":[{"cited_title":"Arbitrary-steps image super-resolution via diffusion inversion,","cited_arxiv_id":null,"evidence_quote":"Provides InvSR, the diffusion-inversion super-resolution model used for category A restoration."},{"cited_title":"Real-esrgan: Training real-world blind super-resolution with pure synthetic data,","cited_arxiv_id":null,"evidence_quote":"Provides Real-ESRGAN, used to restore text regions that the super-resolution model handles poorly."},{"cited_title":"Resolution-robust large mask inpainting with fourier convolutions,","cited_arxiv_id":null,"evidence_quote":"Provides LaMa, the inpainting model that fills masked strong-reflection regions in category B."},{"cited_title":"Simple baselines for image restoration,","cited_arxiv_id":null,"evidence_quote":"Provides NAFNet, used both to refine weak reflections and to perform general motion deblurring."},{"cited_title":"Towards robust blind face restoration with codebook lookup transformer,","cited_arxiv_id":null,"evidence_quote":"Provides CodeFormer, the face-restoration model applied after deblurring for category C."},{"cited_title":"Yolov9: Learning what you want to learn using programmable gradient information,","cited_arxiv_id":null,"evidence_quote":"Provides YOLO, the detector used to locate strong reflection regions before mask generation."},{"cited_title":"You only segment once: Towards real-time panoptic segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides YOSO, the segmentation model used to refine the reflection mask at pixel level."},{"cited_title":"Paddleocr: An open-source op- tical character recognition system,","cited_arxiv_id":null,"evidence_quote":"Detects text regions in super-resolution images so Real-ESRGAN can restore them separately."}],"review_version":1}