{"id":"ec0baa65-e071-46c9-8dc8-2881960feec5","arxiv_id":"2608.11562","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A physics-grounded synthesis pipeline generates paired reflected and clean videos, a two-stage diffusion model removes reflections in one step, and a new benchmark evaluates video dereflection.","lead":"Videos shot through glass often contain reflections that ruin the scene. This paper builds a closed loop: synthesize realistic reflected videos with glass-like effects, train a fast one-step diffusion model to remove the reflections, and introduce the first benchmark for video dereflection.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"S2R-Ref tests only static reflections: identical Ken Burns transforms are applied to both layers, so independent motion and parallax—the defining video challenge—are never evaluated quantitatively.","rationale":"The reader's weakest_assumption focuses on whether the learned renderer G in Eq. (2) preserves the clean video T, which is a valid data-fidelity concern. However, the more load-bearing gap is the evaluation itself: even if G is perfectly faithful, the full-reference benchmark S2R-Ref does not test the defining video challenge of independent reflection motion or viewpoint-dependent parallax. The paper's own Appendix I limitation statement explicitly acknowledges that parallax is not simulated, so both training and evaluation miss this phenomenon. The central claim of 'temporally coherent, state-of-the-art video dereflection' therefore rests on measurements on static-reflection sequences, which is a weaker empirical basis than advertised. This is not an internal inconsistency, but it is a correctness-risk concern: if the method were evaluated on independent-motion sequences, its performance could degrade substantially, and the SOTA claim would not transfer. The proposed concrete test—a paired synthetic benchmark with independent reflection motion—would settle this directly. Since the reader already marked the verdict CONDITIONAL and noted the need for independent-motion validation in the rationale, my concern does not change the verdict; it reinforces the conditionality. I therefore recommend UNCHANGED, with the condition explicitly tied to demonstrating performance on independent-motion reflections.","tokens_in":18089,"tokens_out":7051,"duration_ms":74783,"concrete_test":"Construct a paired synthetic video benchmark with independent reflection motion: take a clean video T and a separate reflection video R with different optical flow (or different camera motion), composite them as I = alpha_t * T + alpha_r * R with an alpha mask, and evaluate S2R-Removal's PSNR, SSIM, and TC on this benchmark. Compare against the S2R-Ref numbers; if the PSNR drop exceeds ~2 dB or TC degrades significantly, the 'temporally coherent video dereflection' claim does not hold for the targeted independent-motion regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that S2R-Removal achieves temporally coherent, state-of-the-art video dereflection, with the intro explicitly motivating the problem by noting that reflections 'move independently from the transmission layer, and interact with camera motion.' Yet the full-reference benchmark S2R-Ref (Section D.1) is constructed from DRR image pairs converted to 10fps static videos, with identical virtual camera motions (Pan/Zoom crop-and-scale) applied to both the reflected and clean videos. Because the same global transform is applied to both layers, the reflection never moves independently of the transmission; the resulting sequences are equivalent to a static reflection overlaid on a scene undergoing a global camera move. Thus the TC and PSNR numbers in Table 1 are measured on a degenerate case that does not exercise the video-specific phenomenon the method is designed for. S2R-Real contains 19 dynamic videos, but only aggregate human scores are reported; no full-reference or per-subset quantitative evaluation is given. Appendix I explicitly concedes that the synthesis pipeline 'does not explicitly simulate the coupled change between camera motion and reflection geometry, such as viewpoint-dependent reflection parallax.' Therefore, the empirical support for the 'temporally coherent video dereflection' claim is missing precisely in the regime where reflections move or parallax occurs, and the state-of-the-art claim on real video is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a closed-loop framework for video reflection removal. S2R-Synthesis generates paired reflected/clean videos by combining a structure-space physics-grounded augmentation (PGA) of reflection lineart conditions with a Wan2.1-based video diffusion renderer trained on FLUX-generated pseudo reflection videos. S2R-Removal adapts a pretrained video diffusion prior via two-stage training: Stage I performs reflection-aware latent adaptation with an intensity map loss derived from |I-T|, and Stage II applies one-step pixel-geometric refinement with reconstruction, SSIM, and depth-consistency losses. The authors introduce S2R-Bench, comprising S2R-Ref (60 full-reference clips built from DRR image pairs with identical Ken Burns motions applied to both layers) and S2R-Real (50 in-the-wild videos with human evaluation). Experiments on S2R-Bench and image benchmarks report state-of-the-art PSNR/SSIM/TC and human scores, with 87.09 ms/frame inference, faster than non-diffusion baselines.","tokens_in":18379,"tokens_out":4402,"duration_ms":50765,"significance":"If validated, the framework addresses a real bottleneck—the lack of scalable, paired, temporally coherent video dereflection data—and demonstrates that a video diffusion prior can be adapted to a restoration task with efficient one-step inference. The paper's strengths include the physically motivated augmentation primitives with explicit derivations (roughness, thickness, reflectance, partial, static), a self-contained synthesis pipeline, a data ablation (Table 2) showing consistent gains from each PGA component, downstream task improvements in segmentation and detection (Tables 10-11), and the first dedicated video reflection removal benchmark. The internal consistency of the ablations and the clear reporting of training configurations are commendable. However, the full-reference benchmark's construction and several unquantified assumptions limit the strength of the central 'temporally coherent, state-of-the-art video dereflection' claim.","major_comments":[{"comment":"The S2R-Ref benchmark applies identical Ken Burns transforms (Pan or Zoom) to both the reflected video and the clean ground truth, so the reflection layer never moves independently of the transmission layer and viewpoint-dependent parallax is absent. This does not exercise the defining video-specific phenomenon that motivates the paper in Section 1 ('reflections vary over time, move independently from the transmission layer, and interact with camera motion'). Appendix I explicitly concedes that the synthesis pipeline does not simulate the coupled change between camera motion and reflection geometry. Consequently, the PSNR/SSIM/TC numbers in Table 1 measure a static reflection overlaid on a globally transformed scene, and the claim of 'temporally coherent state-of-the-art video dereflection' is not supported in the regime where reflections move or parallax occurs. Please either construct an additional full-reference or per-subset benchmark with independent reflection motion, or clearly scope the claims to the tested setting.","section":"Section D.1, Table 1"},{"comment":"No error bars, confidence intervals, or significance tests are reported for any metric. The S2R-Ref subset contains only 60 clips, the human scores in S2R-Real come from only 15 participants, and several cross-benchmark differences are small (e.g., Nature PSNR 27.89 for Ours vs. 28.08 for GenSIRR). Without variance or per-clip statistics, the state-of-the-art claims on both video and image benchmarks are not yet established. Please report standard deviations across clips or across multiple training runs, and inter-rater agreement for the human evaluation.","section":"Tables 1 and 3"},{"comment":"The paper's headline efficiency claim rests on the one-step deterministic denoising update in Eq. (6), yet no ablation compares one-step inference against multi-step sampling in Stage II. The text asserts that Stage I instils a 'meaningful one-step denoising capability' and cites observation in [6], but the paper does not show that the one-step mapping is stable or that it does not sacrifice quality relative to more sampling steps. Please provide an ablation over the number of denoising steps at inference, and, if the one-step design is retained, report the quality/efficiency trade-off explicitly.","section":"Eq. (6), Section 4.2"},{"comment":"The paper claims 'exact paired supervision' for the synthesized pairs (I, T), but this holds only if the learned renderer G preserves the clean video T and adds reflections without otherwise editing it. Since G is trained on FLUX pseudo-reflection videos without an explicit T-preservation constraint, and since the fused condition E_F is clipped in Eq. (14), the renderer could introduce appearance or geometry changes to T that the removal model then learns to reverse. Please quantify how much G modifies T in non-reflection regions (e.g., report a consistency metric between G(T, E_F) and T in areas where the reflection condition is zero), or add a preservation loss during renderer training to substantiate the 'exact paired supervision' claim.","section":"Section 3.2, Eq. (2)"}],"minor_comments":[{"comment":"The sentence 'which no prior benchmark provides' should be qualified to 'no prior benchmark for video reflection removal provides', since image-level reflection removal benchmarks do exist.","section":"Section 2.2"},{"comment":"The appendix mixes two numbering schemes: sections are sometimes labeled 'A1', 'A2', etc. and sometimes 'A.1', 'A.2'. Please standardize the section labels.","section":"Appendix A"},{"comment":"The table caption does not define 'Trad. [4]' until the text in Section 5.3; please explicitly state that it refers to RGB-space layer blending under Eq. (1).","section":"Table 2"},{"comment":"The zoom-out inference mode consistently outperforms zoom-in, but no explanation is given for this asymmetry. Since the removal model is trained on videos with camera motion, the direction dependence is surprising and deserves a brief discussion.","section":"Table 9"},{"comment":"The attention-map visualization is informative, but the paper should clarify whether the 'DiT Map w/ head' is the same as the predicted intensity map M-hat or a separate attention response; the terminology is currently ambiguous.","section":"Figure 10"},{"comment":"The ablation only reports results on S2R-Ref. Reporting the same ablation on an image benchmark or on a subset of S2R-Real with dynamic motion would strengthen the claim that the depth-consistency loss specifically benefits video geometry rather than merely improving static-frame metrics.","section":"Section 5.4, Table 3"}],"recommendation":"major_revision","confidential_remarks":"The paper's appendix is unusually candid about the lack of parallax simulation, which is a point in its favor; however, that same concession undermines the full-reference evaluation of the video-specific contribution. The authors should consider repositioning the contribution as 'efficient diffusion-based video dereflection with a static-reflection synthesis benchmark' while adding a supplementary dynamic benchmark, or they should add real dynamic-video full-reference results if any can be obtained. The license note for FLUX (non-commercial) is worth verifying given the journal setting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear X,\n\nReal systems contribution, serious evaluation gap. The synthesis pipeline and the one-step diffusion removal model are worth the field's attention, but the headline claim—state-of-the-art video dereflection—is not actually tested in the regime that defines the problem.\n\nWhat is new: the first diffusion-based video reflection removal model I'm aware of, the first paired-video benchmark (even if limited), and a closed-loop synthesis/removal framework that combines structure-space physics-grounded augmentation with a video diffusion renderer. The data ablations in Table 2 are the best part: fixing RDNet and varying only the training data shows a clear +1.78 dB gain from the PGA components over planar-only, and the Stage I/II loss ablation cleanly attributes the gains. External image benchmarks and the BDD100K downstream experiments are sensible supporting evidence.\n\nSoft spots, in order: (1) S2R-Ref applies identical Ken Burns crop/zoom transforms to both the reflected and clean videos, so the reflection never moves independently and parallax never appears. That is exactly the phenomenon the intro motivates the paper with. Appendix I concedes the synthesis pipeline does not model viewpoint-dependent reflection parallax. S2R-Real has 19 dynamic clips, but only aggregate human scores are reported, no per-subset numbers. So the empirical support for \"temporally coherent video dereflection\" is missing where reflections actually move. (2) The renderer G in Eq. (2) is trusted to leave T unchanged, but there is no verification; if G edits the transmission, the exact paired supervision is less exact than claimed. (3) The benchmark is small (60 clips) and comparisons have no error bars—minor, but worth stating. (4) No code or data release is mentioned; for a benchmark-driven paper that is a real limitation.\n\nNone of this kills the core contribution. The method is clearly engineered, the experiments are honestly reported, and the limitation statement is candid. But the title-level claim needs softening: the method is state-of-the-art on static-overlay reflection videos, not on videos with moving reflections or parallax.\n\nFor peer review: yes, send it to a serious referee. The right outcome is major revision: add dynamic-reflection validation with independent motion, report S2R-Real per-subset numbers, verify the renderer's transmission preservation, and release code/data. I'd take it to reading group and would cite the synthesis recipe even while I avoid the SOTA claim.","headline":"Real systems contribution with a serious evaluation gap: the benchmark does not test independent reflection motion, so the SOTA video claim needs softening before acceptance.","tokens_in":18879,"tokens_out":3798,"would_cite":true,"duration_ms":41412,"reading_group":"yes","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 video reflection removal can be solved by a closed loop of physics-grounded synthetic pairs, a one-step diffusion removal model, and a dedicated benchmark.","keywords":["video reflection removal","diffusion models","reflection synthesis","physics-grounded augmentation","lineart structure space","one-step denoising","video benchmark","temporal coherence"],"falsifier":"Take a known clean video, synthesize reflections with S2R-Synthesis at the extremes of the PGA parameter ranges (largest roughness, largest thickness offset, and strongest reflectance), then decode the renderer's output and measure the pixel residual against the clean video in regions the reflection mask leaves untouched: if that residual is not explained by the encoded reflection layer, or if a real-glass test with a known transmission layer shows the one-step model's output drifting from the true transmission outside the training parameter range, the central synthesis-to-removal claim fails.","tokens_in":17893,"feed_emoji":"🪟","tokens_out":8610,"duration_ms":85024,"temperature":0.7,"pith_summary":"This paper tries to establish that video reflection removal can be solved end-to-end as a closed loop: a physics-grounded synthesis pipeline generates large-scale, pixel-aligned reflected/clean video pairs; a diffusion-based removal model learns to undo the reflections; and a purpose-built benchmark measures fidelity and human perception. The synthesis side replaces naive RGB blending with compositing of lineart structure conditions through a learned video renderer, under clip-level control of six glass effects: roughness, thickness, reflectance, partial coverage, static sources, and planar geometry. The removal side adapts a pretrained video diffusion model in two stages, first teaching reflection-aware latent adaptation from a residual-derived intensity map and then one-step pixel-geometric refinement, so the final model recovers the clean transmission in a single denoising step at 87.09 ms per frame, faster than non-diffusion baselines. A sympathetic reader would care because the paper targets the two known bottlenecks—missing paired video data and temporal coherence—and claims to remove both.","feed_headline":"One diffusion step removes video reflections at 87 ms per frame","feed_subtitle":"Physics-grounded synthetic pairs train the first diffusion video dereflection model; a new benchmark tracks it.","key_machinery":"The load-bearing object is the structure-space synthesis identity that replaces the linear layer model: $E_F = \\operatorname{clip}(E_T + \\mathcal{A}(E_R;\\theta_g), 0, 1)$ and $I = G(T, E_F)$, where $\\mathcal{A}$ is Physics-Grounded Augmentation (PGA)—roughness as Gaussian blur, thickness as a shifted weighted copy, reflectance as affine modulation, plus partial, static, and planar modes—and $G$ is a learned video renderer. This identity turns physical glass parameters into clip-level structural controls, so synthesized pairs stay temporally coherent. On the removal side, the engine is the two-stage objective $\\mathcal{L}_{\\mathrm{stage1}} = \\lambda_{\\mathrm{diff}}\\mathcal{L}_{\\mathrm{diff}} + \\lambda_{\\mathrm{int}}\\mathcal{L}_{\\mathrm{int}}$ followed by $\\mathcal{L}_{\\mathrm{stage2}} = \\lambda_{\\mathrm{rec}}\\mathcal{L}_{\\mathrm{rec}} + \\lambda_{\\mathrm{ssim}}\\mathcal{L}_{\\mathrm{ssim}} + \\lambda_{\\mathrm{dep}}\\mathcal{L}_{\\mathrm{depth}}$, with the emergent one-step denoising update $\\hat{z}_0 = D^{(1)}_\\theta(z_\\tau, \\tau, z_I)$ converting latent diffusion into a fast, geometry-anchored restoration.","core_discovery":"This paper claims that a video diffusion prior can be turned into a practical reflection remover if the training pipeline is closed-loop. S2R-Synthesis generates paired reflected/clean videos by fusing transmission and reflection lineart conditions through a learned renderer, with Physics-Grounded Augmentation controlling roughness, thickness, reflectance, partial coverage, static sources, and planar geometry; S2R-Removal then adapts a pretrained video diffusion backbone in two stages—latent adaptation with residual-derived reflection-intensity supervision, followed by one-step pixel-geometric refinement using reconstruction, SSIM, and depth-consistency losses. The result, the paper argues, is the first diffusion-based video reflection removal model, recovering the clean transmission in a single denoising step, reaching 28.84 dB PSNR on the new S2R-Ref benchmark and 0.787/0.980 human-scored removal/preservation on real-world videos, while running at 87.09 ms per frame.","pith_inferences":["The PGA parameterization suggests a modular recipe for other degradations: the same structure-space augmentation with a learned renderer could synthesize rain, snow, dirt, or glare overlays, and the two-stage removal objective would transfer wherever the target is a preserved underlying layer.","The reflection-intensity head effectively produces a dense reflection-strength map as a byproduct; this could serve as weak supervision for reflection segmentation or as an interpretability signal for where the diffusion model decides to edit.","If the emergent one-step capability is a general property of latent diffusion adaptation, then many video restoration tasks currently confined to slow multi-step diffusion inference could adopt the same two-stage latent-adaptation-then-pixel-geometric-refinement schedule."],"forward_implications":["Paired video reflection data stops being a bottleneck: the S2R-Synthesis pipeline can generate large, aligned, and physically varied reflected/clean pairs on demand, including effects that are difficult to capture in reality.","Diffusion-based dereflection becomes practical for video: because inference is one step, the model runs at 87.09 ms per frame and is about 1.67 times faster than the next-best baseline despite using a diffusion prior.","Standard image benchmarks no longer set the ceiling: the video-trained model also reaches state-of-the-art PSNR and SSIM on public image benchmarks through zoom-based pseudo-video inference.","Dereflection pays off downstream: driving-area segmentation and vehicle detection on reflection-contaminated driving videos improve after removal, suggesting preprocessing gains for perception systems.","A common evaluation protocol becomes possible: S2R-Bench supplies 60 paired full-reference videos and 50 real-world videos with human scoring, filling the missing benchmark gap for the task."],"supporting_citations":[{"why":"Supplies the Real image benchmark used for cross-domain evaluation of the removal model.","marker":"[3]"},{"why":"Defines the linear layer-formation model underlying Eq. (1) and the traditional RGB-blending synthesis baseline compared in the data ablations.","marker":"[4]"},{"why":"Provides a diffusion-prior replacement baseline and the paired image data from which the S2R-Ref subset is built, with that baseline excluded to avoid data leakage.","marker":"[5]"},{"why":"Contributes the observation that latent diffusion training yields an emergent one-step denoising capability, which Stage II of S2R-Removal exploits.","marker":"[6]"},{"why":"Generates the per-frame pseudo reflection videos used only as supervision when training the synthesis renderer.","marker":"[9]"},{"why":"Supplies the pretrained video diffusion backbone on which both the renderer and the removal model are built.","marker":"[10]"},{"why":"Provides the non-diffusion image baseline and the fixed backbone used to isolate the effect of different synthetic training data in the synthesis ablations.","marker":"[18]"},{"why":"Supplies the SIR2 image benchmark used for cross-domain evaluation.","marker":"[22]"},{"why":"Provides the frozen depth estimator used in the mask-weighted depth-consistency loss that anchors the one-step output geometrically.","marker":"[33]"}],"fun_headline_variants":["Diffusion video dereflection in one step, 87 ms per frame","Physics-grounded synthesis powers first diffusion video dereflection","Closed-loop synthesis and diffusion remove video reflections","First diffusion video dereflection model, one-step removal","Video dereflection at 87 ms with physics-grounded training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole loop depends on the renderer $G$ adding reflections without changing the clean transmission video: if the lineart fusion or the learned renderer alters $T$ rather than layering a reflection on top, the removal model trains on artifacts and the benchmark gains may not transfer to real glass.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion video dereflection in one step, 87 ms per frame","Physics-grounded synthesis powers first diffusion video dereflection","Closed-loop synthesis and diffusion remove video reflections","First diffusion video dereflection model, one-step removal","Video dereflection at 87 ms with physics-grounded training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000923,"raw_usage":{"total_tokens":3988,"prompt_tokens":1008,"completion_tokens":2980,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":624,"completion_tokens_details":{"reasoning_tokens":2896}},"tokens_in":624,"tokens_out":2980,"duration_ms":21116,"temperature":1.0,"reasoning_tokens":2896,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:36:12.712599+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a known clean video, synthesize reflections with S2R-Synthesis at the extremes of the PGA parameter ranges (largest roughness, largest thickness offset, and strongest reflectance), then decode the renderer's output and measure the pixel residual against the clean video in regions the reflection mask leaves untouched: if that residual is not explained by the encoded reflection layer, or if a real-glass test with a known transmission layer shows the one-step model's output drifting from the true transmission outside the training parameter range, the central synthesis-to-removal claim fails.","supporting_citations":[{"cited_title":"Single image reflection separation with perceptual losses,","cited_arxiv_id":null,"evidence_quote":"Supplies the Real image benchmark used for cross-domain evaluation of the removal model."},{"cited_title":"Single image reflection separation via component synergy,","cited_arxiv_id":null,"evidence_quote":"Defines the linear layer-formation model underlying Eq. (1) and the traditional RGB-blending synthesis baseline compared in the data ablations."},{"cited_title":"Dereflection any image with diffusion priors and diversified data,","cited_arxiv_id":null,"evidence_quote":"Provides a diffusion-prior replacement baseline and the paired image data from which the S2R-Ref subset is built, with that baseline excluded to avoid data leakage."},{"cited_title":"FLUX.2: Frontier Visual Intelligence,","cited_arxiv_id":null,"evidence_quote":"Generates the per-frame pseudo reflection videos used only as supervision when training the synthesis renderer."},{"cited_title":"Reversible decoupling network for single image reflection removal,","cited_arxiv_id":null,"evidence_quote":"Provides the non-diffusion image baseline and the fixed backbone used to isolate the effect of different synthetic training data in the synthesis ablations."},{"cited_title":"Benchmarking single-image reflectionremovalalgorithms,","cited_arxiv_id":null,"evidence_quote":"Supplies the SIR2 image benchmark used for cross-domain evaluation."},{"cited_title":"Learning to recover 3d scene shape from a single image,","cited_arxiv_id":null,"evidence_quote":"Provides the frozen depth estimator used in the mask-weighted depth-consistency loss that anchors the one-step output geometrically."}],"review_version":1}