{"id":"4426dc99-44c3-4cc2-a870-0acbd02fac2e","arxiv_id":"2607.05261","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":13,"one_line_summary":"FlowMark learns content-adaptive spatial masks for video watermark embedding, achieving 50+ dB PSNR, 128-bit capacity, and robustness to compression, temporal edits, and social media pipelines.","lead":"FlowMark is a video watermarking system that automatically learns where to embed invisible watermarks in each frame, achieving high visual quality and robustness to compression, temporal edits, and social media re-encoding. A smart generalist might read it to understand how AI can help verify the authenticity and provenance of video content in an era of synthetic media.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The 'flicker elimination' claim — FlowMark's primary differentiator from VideoSeal — lacks any quantitative temporal flicker metric; it is supported only by L1 residual norms, VMAF (a general quality metric), and qualitative visualizations.","rationale":"The reader correctly identified the most load-bearing concern. FlowMark's primary contribution over VideoSeal is flicker-free temporal consistency with per-frame unique watermarks (as opposed to VideoSeal's repeated-message approach). If this claim is not properly substantiated, the paper's main differentiator is unsupported. The concern is not that the method necessarily fails — the temporal consistency loss and visualizations suggest it likely helps — but that the evaluation does not include a metric that directly measures the claimed property. This is a measurement gap, not an internal inconsistency or a disagreement with consensus. The quantitative results on PSNR, bit accuracy, robustness, and tamper detection (Tables 1–2, Fig. 5) are well-supported and appear sound. The ablation study (Table 3) is thorough and honestly reports trade-offs (e.g., dark-region masking improves PSNR but slightly reduces robustness). The full-resolution generalization claim is also only qualitatively supported, but this is a secondary claim. The social media robustness test (20 videos) has a small sample size, but the results are directionally consistent with the compression robustness results. No code or weights are released, which limits independent verification, but this is a reproducibility concern rather than a correctness concern. The verdict should remain CONDITIONAL — the paper presents a legitimate, well-engineered system with strong quantitative results on standard metrics, but the headline 'flicker elimination' claim requires a dedicated temporal flicker metric to be fully substantiated.","tokens_in":19173,"tokens_out":2991,"duration_ms":70801,"concrete_test":"Compute a dedicated temporal flicker metric on the SA-V test set (155 videos) for FlowMark and all baselines. Specifically, compute the per-pixel temporal variance of the residual R_t = I_wm^t - I^t across consecutive frame windows (e.g., sliding window of 5 frames), then aggregate spatially (mean or 95th percentile). Additionally, compute temporal SSIM variance between consecutive watermarked frames vs. consecutive original frames. If FlowMark's temporal residual variance or temporal SSIM variance is not significantly lower than VideoSeal's, the 'flicker elimination' claim weakens. This requires only the already-generated watermarked outputs and no retraining.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central differentiating claim of FlowMark over prior work is 'flicker-free temporal consistency' (abstract, §1, §4.2). This claim rests on the Temporal Change MSE loss (Eq. 8), which minimizes ||ΔI_wm^t - ΔI^t||² between consecutive frames. However, the evaluation of whether this actually eliminates perceptual flicker is insufficient:\n\n1. The L1 residual norm (||R||₁ = 0.005, Fig. 1, Fig. 4) measures embedding magnitude, not temporal variation. A method could have low residual norm yet still flicker if the residual pattern shifts spatially across frames.\n\n2. VMAF (98.95) is a general video quality metric that correlates with human judgment of overall quality, not a flicker-specific measure. It does not isolate temporal instability from spatial quality.\n\n3. The secret-invariance comparison (Fig. 9) is conducted on SA-I images, not video frames, so it demonstrates spatial stability of residuals across secrets but not temporal flicker in actual video playback.\n\n4. The ablation (Table 3) shows removing Ltcm drops PSNR to 37.19 dB and robust accuracy to 0.79, but neither metric directly measures flicker. A model could have high PSNR yet still exhibit perceptible temporal flicker.\n\n5. The qualitative visualizations (Figs. 11–16) show residual maps but do not constitute a controlled perceptual study or objective flicker measurement.\n\nThe concern is that Eq. 8 constrains the difference of differences (second-order temporal derivative), which is a necessary but not sufficient condition for perceptual flicker elimination. Flicker perception depends on spatial frequency content of temporal changes, local contrast, and viewing conditions — none of which are captured by the L2 penalty in Eq. 8 or the reported metrics. If the temporal consistency loss is insufficient for high-motion scenes, scene cuts, or texture-changing regions not well-represented in the 155-video SA-V test set, flicker could persist undetected by the current evaluation protocol.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"FlowMark proposes a mask-guided video watermarking framework that learns spatial embedding regions via a Mask Predictor network, rather than requiring user-supplied masks. The system combines dark-region adaptive residual modulation, a temporal change consistency loss, and a three-stage curriculum training strategy (clean → image augmentations → video codec distortions). A structured 128-bit per-frame embedding (112-bit video ID + 16-bit frame index) enables temporal tamper detection (frame swap, insertion, deletion, resampling, interpolation). Experiments compare against 7 baselines on SA-1B (image) and SA-V (video) datasets across valuemetric, geometric, compression, and combined distortions, plus social media pipeline tests (YouTube, Facebook), temporal edit detection, adversarial attack resilience, and 3 ablation studies. FlowMark reports up to 50.08 dB PSNR, perfect bit accuracy under identity conditions, 100% temporal tamper detection, and VMAF of 98.95 on SA-V.","tokens_in":19541,"tokens_out":1703,"duration_ms":79154,"significance":"The paper addresses a practically important problem: achieving per-frame unique watermarks in video without introducing perceptual flicker, which is a known limitation of frame-independent embedding. The learned mask predictor removes the need for manual mask selection required by prior region-based methods (WAM, MaskWM), which is a meaningful practical contribution. The structured bit embedding for temporal provenance (frame-level traceability) is a useful design choice for content authenticity applications. The experimental evaluation is extensive in scope, covering multiple distortion categories, real-world social media pipelines, and temporal edit scenarios. The three-stage curriculum and the dark-region adaptive modulation are reasonable engineering contributions. However, the central differentiating claim of 'flicker-free temporal consistency' lacks a dedicated quantitative metric, which weakens the evidentiary support for the paper's primary contribution claim.","major_comments":[{"comment":"§4.2, Table 1, and §1 (abstract): The claim that FlowMark 'effectively eliminates perceptual flicker' (abstract) and achieves 'flicker-free temporal consistency' (§4.2) is the primary differentiator from VideoSeal, yet no quantitative temporal flicker metric is reported. The L1 residual norm (||R||_1 = 0.005, Fig. 4) measures embedding magnitude, not temporal variation — a method could have low residual norm yet still flicker if the residual pattern shifts spatially across frames. VMAF (98.95) is a general video quality metric, not a flicker-specific measure. The secret-invariance comparison (Fig. 9) is conducted on SA-I images, not video frames, so it demonstrates spatial stability across secrets but not temporal flicker during video playback. The ablation (Table 3) shows removing L_tcm drops PSNR to 37.19 dB and robust accuracy to 0.79, but neither metric directly measures flicker. The","section":null},{"comment":"claim would be substantially strengthened by reporting a temporal stability metric such as temporal SSIM variance, a perceptual flicker index, or frame-to-frame residual difference norms on actual video sequences. Without such a metric, the 'eliminates perceptual flicker' claim is supported only by qualitative visualizations and proxy metrics.","section":null},{"comment":"§3.2 and §4.3: The structured bit embedding (b_vid || b_frm with 16-bit frame index) supports at most 65,536 unique frame indices. The paper does not discuss what happens for videos exceeding this length, nor whether the frame index wraps or resets. For practical deployment in long-form video or continuous streaming, this limitation should be acknowledged and the handling of overflow specified.","section":null},{"comment":"Table 3 (ablation): Removing the dark mask improves robust accuracy from 0.80 to 0.86 while reducing PSNR from 50.41 to 40.18 dB. The paper frames this as a fidelity-robustness trade-off, but the robustness improvement (0.80 → 0.86) is non-trivial and suggests the dark mask may be overly conservative. The paper should discuss whether a less aggressive dark mask (e.g., higher d_min or τ) could achieve a better operating point, or provide justification for the chosen parameter values beyond the single setting reported.","section":null},{"comment":"§4.3 (social media robustness): The social media robustness test uploads 20 videos to YouTube and Facebook, but no quantitative bit accuracy numbers are reported in the main text — only a qualitative statement that 'FlowMark maintains high bit accuracy.' Figure 5(b) appears to show this result, but the exact bit accuracy values under YouTube and Facebook compression are not tabulated. Given that social media robustness is a headline claim (Fig. 1b), the specific numbers should be reported in a table or explicitly in the text.","section":null}],"minor_comments":[{"comment":"§3, Eq. (2): The dark-region mask D(x,y) uses a sigmoid with argument (L(x,y) - τ)^s. When L < τ, (L - τ)^s may be problematic for non-integer s (here s=10 is fine, but the formulation should clarify the sign convention or use an absolute value to avoid ambiguity for general s).","section":null},{"comment":"Table 1: MaskWM is cited as [30] in the table but [31] in the related work text. Please reconcile.","section":null},{"comment":"Table 1: FlowMark's mask accuracy (0.95 on SA-1b, 0.96 on SA-V) is lower than MaskWM (1.00 on both). This is not discussed. A brief note on why FlowMark's mask accuracy is slightly lower would be helpful.","section":null},{"comment":"§4.1: The Adobe Stock video dataset used for training is not described (size, resolution, duration). Since training data composition can affect generalization, a brief description would be appropriate.","section":null},{"comment":"Fig. 9 caption: 'SA-I images' should likely be 'SA-1B images' for consistency with the rest of the paper.","section":null},{"comment":"§6 (Full Resolution Watermarking): The claim that upscaling a 256×256 learned residual to full resolution preserves robustness and perceptual quality is stated without quantitative evaluation at multiple resolutions. Fig. 10 shows qualitative examples but no PSNR/SSIM/bit-accuracy numbers at native resolution. Adding even a small quantitative table would strengthen this section.","section":null},{"comment":"References [23] and [24] both cite VideoSeal but with slightly different formatting (one as 'Video seal', one as 'VideoSeal'). Please consolidate.","section":null},{"comment":"Table 4: The caption mentions 'following VideoSeal [23] evaluation' for GFLOPs, but the reference number should be checked for consistency (VideoSeal is cited as [16] in §3 and [23] elsewhere).","section":null},{"comment":"§4.4: The watermark strength ablation (Fig. 8) uses parameter α, but α is not defined in the method section. It should be clarified whether α corresponds to µ in Eq. (3) or is a separate inference-time parameter.","section":null}],"recommendation":"major_revision","confidential_remarks":"The core method is sound and the experimental breadth is commendable. The primary concern is the gap between the 'flicker elimination' claim and the evidence presented — this is the paper's main selling point over VideoSeal and it needs quantitative support. Adding a temporal flicker metric on video sequences would likely resolve this concern and could be done without new training. The social media numbers also need to be explicitly reported. I do not see fundamental issues with the method design; the concerns are about evaluation completeness for the central claims."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a thorough and constructive report. The core concern—lack of a dedicated quantitative flicker metric—is well-taken, and we will address it with new experiments. The remaining comments identify legitimate gaps in discussion and reporting that we will incorporate in revision.","responses":[{"response":"The referee is correct that none of our currently reported metrics directly isolate temporal flicker. The L1 residual norm, VMAF, and secret-invariance analysis are all proxy measures that do not specifically capture frame-to-frame spatial variation of the watermark residual during video playback. We agree this is a genuine gap given that flicker-free temporal consistency is our primary differentiating claim. In revision, we will add a dedicated quantitative flicker metric computed on actual SA-V video sequences. Specifically, we will report frame-to-frame residual difference norms (||R_t - R_{t-1}||_1) and temporal SSIM variance across consecutive watermarked frames for FlowMark and all baselines. This will directly measure whether the residual pattern shifts spatially across frames, which is the mechanism by which frame-independent embedding produces visible flicker. We will also soften the absolute language in the abstract from 'effectively eliminates perceptual flicker' to a more precise claim grounded in the new metric, e.g., 'significantly reduces temporal flicker as measured by [metric].'","revision_made":"yes","referee_comment":"The claim that FlowMark 'effectively eliminates perceptual flicker' and achieves 'flicker-free temporal consistency' is the primary differentiator from VideoSeal, yet no quantitative temporal flicker metric is reported. The L1 residual norm measures embedding magnitude, not temporal variation. VMAF is a general video quality metric, not a flicker-specific measure. The secret-invariance comparison is on SA-I images, not video frames. The ablation does not directly measure flicker. The claim would be substantially strengthened by reporting a temporal stability metric such as temporal SSIM variance, a perceptual flicker index, or frame-to-frame residual difference norms on actual video sequences."},{"response":"This is a valid observation that we failed to address. The 16-bit frame index does limit unique frame indexing to 65,536 frames (approximately 45 minutes at 24fps). For longer content, the index would wrap, which could create ambiguity in frame ordering for tamper detection. We will add a paragraph to Section 3.2 acknowledging this limitation and describing the trade-off: allocating more bits to the frame index (e.g., 24 bits, supporting ~18 hours at 24fps) reduces the video ID capacity correspondingly. We will note that for the vast majority of practical use cases—social media clips, short-form video, and typical streaming segments—the 16-bit index is sufficient, and that the allocation is configurable based on deployment requirements. We will also note that for continuous streaming, a segment-based approach with periodic video ID rotation can be used.","revision_made":"yes","referee_comment":"The structured bit embedding (b_vid || b_frm with 16-bit frame index) supports at most 65,536 unique frame indices. The paper does not discuss what happens for videos exceeding this length, nor whether the frame index wraps or resets. For practical deployment in long-form video or continuous streaming, this limitation should be acknowledged and the handling of overflow specified."},{"response":"The referee raises a fair point. The 6-point robustness gain from removing the dark mask is indeed non-trivial, and our current treatment does not adequately justify the chosen parameters (d_min=0.2, tau=0.3, s=10) or explore intermediate settings. We will address this in two ways. First, we will add discussion acknowledging that the dark mask represents a deliberate fidelity-first operating point chosen for provenance applications where visual quality is paramount and the 0.80 robust accuracy is already strong. Second, we will add an ablation over d_min values (e.g., d_min in {0.0, 0.1, 0.2, 0.4, 0.6}) to show the trade-off curve and demonstrate that intermediate settings exist. We note that d_min=0.0 effectively disables the dark mask (matching the current ablation row), so this sweep will characterize the full spectrum the referee requests.","revision_made":"partial","referee_comment":"Removing the dark mask improves robust accuracy from 0.80 to 0.86 while reducing PSNR from 50.41 to 40.18 dB. The paper frames this as a fidelity-robustness trade-off, but the robustness improvement (0.80 to 0.86) is non-trivial and suggests the dark mask may be overly conservative. The paper should discuss whether a less aggressive dark mask could achieve a better operating point, or provide justification for the chosen parameter values beyond the single setting reported."},{"response":"The referee is correct. For a headline claim, reporting only a qualitative statement without explicit numerical values is insufficient. We will add a small table reporting bit accuracy for FlowMark under YouTube and Facebook compression separately, alongside the number of videos tested and the corresponding PSNR after platform re-encoding. If space permits in the main text, this will be a dedicated table; otherwise it will appear as explicit numerical values in the text of Section 4.3 with the full table in the supplementary material.","revision_made":"yes","referee_comment":"The social media robustness test uploads 20 videos to YouTube and Facebook, but no quantitative bit accuracy numbers are reported in the main text—only a qualitative statement that 'FlowMark maintains high bit accuracy.' Figure 5(b) appears to show this result, but the exact bit accuracy values under YouTube and Facebook compression are not tabulated. Given that social media robustness is a headline claim, the specific numbers should be reported in a table or explicitly in the text."}],"tokens_in":19447,"tokens_out":1218,"duration_ms":60181,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"Two things to know up front: FlowMark is a well-engineered video watermarking system that combines several known techniques into a practical whole, and its central differentiating claim — flicker-free temporal consistency — is supported by strong quantitative results on fidelity and robustness but only weakly supported evidence on the flicker question itself. The reader's stress-test concern about the missing flicker metric is largely correct and is the main soft spot in an otherwise solid paper. What's genuinely new: a learned mask predictor that automatically selects embedding regions (prior work like WAM and MaskWM require user-supplied masks), a structured per-frame bit scheme encoding both video ID and frame index for temporal tamper detection, and a three-stage curriculum that moves from clean training to codec-level distortions. The structured bit embedding is a clean idea — it gives you frame-level provenance that VideoSeal's repeated-message approach can't offer. The experimental evaluation is extensive: 7 baselines, 2 datasets, 4 distortion categories, social media pipeline tests, and 3 ablation studies. The numbers are strong — 50.41 dB PSNR on SA-V, perfect bit accuracy under identity, 100% temporal tamper detection. The ablation table (Table 3) does useful work showing each component contributes. The soft spot is real and proportionate to how much the paper leans on it. The abstract and introduction frame flicker elimination as the primary advantage over VideoSeal, but the evaluation of this claim is insufficient. The L1 residual norm measures embedding magnitude, not temporal variation. VMAF is a general quality metric, not flicker-specific. The secret-invariance comparison (Fig. 9) is on still images, not video. The temporal change MSE loss (Eq. 8) is a second-order temporal smoothness penalty — necessary but not sufficient for perceptual flicker elimination. None of this means the method doesn't work; the visualizations are suggestive and the loss design is reasonable. But the paper overclaims relative to its evidence here. A temporal SSIM variance, a perceptual flicker index, or even a small user study would close this gap. Minor concerns: the social media test uses only 20 videos, the adversarial evaluation requires fine-tuning on the attack type, and no code or weights are released. The 13 free parameters are a lot but the ablations show sensitivity is manageable. This paper is for researchers and engineers working on content provenance, video authentication, and watermarking systems. It deserves a serious referee — the core contribution is real and the experimental work is substantial. The referee should push hard on the flicker metric gap and request either a quantitative temporal stability measure or a toning down of the 'eliminates flicker' language to match what's actually demonstrated.","headline":"Solid engineering advance in video watermarking; the 'flicker elimination' claim needs a real temporal metric.","tokens_in":20169,"tokens_out":639,"would_cite":true,"duration_ms":48466,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Learned mask placement eliminates flicker in video watermarks","keywords":["video watermarking","mask-guided embedding","temporal consistency","content-adaptive watermarking","media provenance","tamper detection","perceptual quality"],"falsifier":"A video sequence with rapid, irregular texture changes (e.g., fireworks, fast camera motion through foliage) where the temporal change consistency loss cannot keep the watermark residual co-moving with natural motion, producing measurable flicker under a temporal SSIM variance metric.","tokens_in":19198,"feed_emoji":"🎬","tokens_out":1374,"duration_ms":61221,"temperature":0.7,"pith_summary":"FlowMark is a video watermarking system that learns, frame by frame, where in the image to hide a 128-bit watermark signal. Instead of spreading the watermark uniformly across every pixel or relying on a human to draw a region mask, a small neural network predicts which spatial regions of each frame are most perceptually tolerant of added signal — typically high-texture or motion-stable areas — and confines the watermark there. The watermark residual is further attenuated in dark regions where compression artifacts would make it visible. A temporal change consistency loss forces the watermark to move with the video's natural motion rather than jittering independently from frame to frame, which is the primary cause of the flicker that plagues prior frame-by-frame watermarking methods. The paper also introduces a structured bit layout: 112 bits encode a persistent video identifier and 16 bits encode a frame index, so each frame carries a unique, traceable signature without sacrificing temporal smoothness. Trained through a three-stage curriculum that escalates from clean frames to image-level distortions to real video codec compression (H.264, H.265, VP9, AV1), FlowMark achieves perfect bit recovery under identity conditions, up to 50.41 dB PSNR on video, 100% detection of temporal tampering (frame swap, insertion, deletion, resampling, interpolation), and competitive robustness after YouTube and Facebook re-encoding, outperforming all tested baselines on the SA-V video benchmark.","feed_headline":"Learned mask placement eliminates flicker in video watermarks","feed_subtitle":"A neural mask predictor confines 128-bit watermarks to perceptually tolerant regions of each frame, achieving 50 dB PSNR and 100% tamper检测.","key_machinery":"Mask Predictor (straight-through-estimated binary spatial mask), dark-region adaptive residual modulation (BT.601 luminance thresholding), temporal change consistency loss (Eq. 8), structured 128-bit embedding (112-bit video ID + 16-bit frame index), three-stage training curriculum (clean → image distortions → video codec compression)","core_discovery":"The central mechanism is the combination of a learned spatial mask predictor with a temporal change consistency objective. The mask predictor outputs a binary embedding map via a straight-through estimator, localizing watermark energy in perceptually resilient regions. The temporal change consistency loss (Eq. 8) minimizes the discrepancy between the frame-to-frame difference of watermarked frames and the frame-to-frame difference of original frames, forcing the watermark residual to co-move with natural video dynamics rather than oscillate independently. Together with luminance-adaptive dark-region attenuation and a residual total variation regularizer, this produces watermark residuals the","pith_inferences":["The claim of 'eliminating perceptual flicker' rests on qualitative visualizations and L1 residual norms rather than a dedicated temporal flicker metric; high-motion scenes with rapid texture changes could expose cases where the temporal consistency loss is insufficient.","The 128-bit capacity ceiling (112 video ID + 16 frame index) limits frame indices to 65,536 frames — roughly 45 minutes at 24 fps — which may constrain long-form video provenance without re-keying or segmented identifiers.","The mask predictor's ratio regularization toward 50% coverage may be suboptimal for scenes where the perceptually resilient region is much smaller or larger than half the frame; an adaptive ratio target could improve quality in extreme content.","Social media robustness was tested on 20 videos from a validation set; broader platform coverage (TikTok, Instagram, WhatsApp) with their specific compression pipelines would strengthen the real-world deployment claim."],"forward_implications":["Per-frame unique watermarks with temporal coherence enable fine-grained provenance: each frame in a video can be independently authenticated and ordered, supporting forensic analysis of manipulated video.","The learned mask predictor could be extended to prioritize embedding in semantically meaningful regions (e.g., faces, text) for targeted provenance or to avoid sensitive regions entirely.","If the structured bit scheme scales beyond 16-bit frame indices, it could support longer videos or hierarchical provenance chains encoding editing history within the watermark itself.","The three-stage curriculum strategy (clean → augmentation → codec) may generalize to other media watermarking domains where distortion severity must be introduced gradually."],"fun_headline_variants":["Learned masks embed 128-bit watermarks without perceptual flicker","Automatic region selection yields flicker-free video watermarks","Mask-guided embedding survives temporal edits and social media re-encoding","FlowMark couples learned masks with temporal consistency to suppress flicker"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The paper asserts that flicker is eliminated, but the evidence for temporal stability is primarily qualitative (residual visualizations, difference maps) or proxied by L1 residual norm. No formal flicker metric — such as temporal variance of SSIM or a perceptual flicker index — is reported across frames, so the claim may not hold for high-motion or texture-changing scenes outside the SA-V test distribution.","fun_headline_variants_meta":{"raw":{"variants":["Learned masks embed 128-bit watermarks without perceptual flicker","Automatic region selection yields flicker-free video watermarks","Mask-guided embedding survives temporal edits and social media re-encoding","FlowMark couples learned masks with temporal consistency to suppress flicker"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":597,"prompt_tokens":526,"completion_tokens":71,"prompt_tokens_details":null},"tokens_in":526,"tokens_out":71,"duration_ms":25214,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-07T20:48:18.025462+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"A video sequence with rapid, irregular texture changes (e.g., fireworks, fast camera motion through foliage) where the temporal change consistency loss cannot keep the watermark residual co-moving with natural motion, producing measurable flicker under a temporal SSIM variance metric.","supporting_citations":[],"review_version":1}