{"id":"98cd1a18-12fa-43f7-a286-4fd951290347","arxiv_id":"2509.06803","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"New 1000 FPS benchmarks, MIORe and VAR-MIORe, provide controlled and extreme motion blur to challenge deblurring, frame interpolation, and optical flow models.","lead":"MIORe and VAR-MIORe are two new video datasets captured at 1000 FPS that generate motion blur by averaging sharp frames according to how fast the scene moves. They add an explicit dial for motion strength, from mild to extreme, to stress-test deblurring, frame interpolation, and optical flow algorithms.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Adaptive blur-control rule in Sec 4.3 is unspecified and mathematically inconsistent with the claimed 30-pixel target for slow scenes; 'consistent and controlled motion blur' unverified.","rationale":"Read in good faith: The paper proposes a plausible high-speed multi-task benchmark. The strongest claim, the VAR-MIORe motion-amplitude control, is supported by the frame-count sweep in Table 5 and Fig. 1; for VAR-MIORe the amplitude is indeed explicit in the offsets. My concern is the MIORe construction rule that supposedly homogenizes blur. The math shows the described rule cannot achieve the 30-px target for subpixel motion, and the paper gives no exact mapping, so the key mechanism is under-specified and possibly inconsistent. This is not a disagreement with field consensus; it is an internal correctness risk. The paper deserves credit for using 1000 FPS capture, multiple lenses, and for flagging the OF labels as pseudo-GT; those aspects are not the attack. If the data/code release resolves the mapping and the total-flow check passes, the paper's central claims are likely sound. If not, the MIORe deblurring benchmark would still be useful but its core controlled-blur novelty would be weakened. Hence I do not move the reader's CONDITIONAL verdict; the concern is a concrete condition to fulfill, not a rejection.","tokens_in":12873,"tokens_out":10333,"duration_ms":107280,"concrete_test":"Obtain the raw sequences and the exact averaging rule (formula/pseudocode) from the project page; for a stratified sample of sequences (e.g., 5 from each Fig. 8 bin), compute per-frame optical flow with an independent estimator (RAFT), derive N from the paper's mapping, and measure the actual total flow between the left and right sharp frames of each synthesized blur. Check whether the total-flow distribution is centered at ~30 px with <2x variation across bins; if not, the 'consistent and controlled motion blur' claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that MIORe 'generates consistent motion blur' via adaptive frame averaging rests on the rule in Sec 4.3, but the rule is not stated as a function. The paper says it targets 'a mid-level benchmark of 30 pixels of optical flow' and illustrates: high-flow scenes (about 10 px/frame) use 3 frames; subpixel scenes use up to 30 frames. If total blur magnitude is roughly per-frame flow v times number of averaged frames N, then holding the total at 30 px requires N = 30/v. The examples match at v=10 (N=3) and v=1 (N=30), but for v<1 the required N exceeds the cap (e.g., v=0.5 needs 60 frames, while the text caps at 30 and Fig. 8 caps at 35). The resulting total flow would be ~15 px, not 30, so slow scenes would have systematically lighter blur than fast scenes unless an additional criterion is used. No formula, pseudocode, or flow thresholds for the six Fig. 8 bins are given. Because the entire deblurring benchmark is synthesized by this rule, the reproducibility and the claimed homogeneity of blur magnitude are not established. The OF circularity noted by the reader is real but secondary: the OF task is already labeled 'pseudo-ground truth' in Sec 5.3.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces two multi-task video datasets, MIORe and VAR-MIORe, built from 1000 FPS 1920x1080 recordings of 333 in-the-wild sequences captured with professional prime/zoom lenses. Blurry images are synthesized by averaging varying numbers of high-speed frames, with the number of frames selected from computed optical-flow statistics; the left, middle, and right sharp frames are kept as inputs/ground truth for video frame interpolation, motion deblurring, and optical flow estimation. VAR-MIORe extends MIORe to a continuous range of integration offsets (1-249 ms) to explicitly vary motion amplitude. The paper benchmarks several state-of-the-art deblurring, VFI, and OF methods and reports results by motion-intensity splits and integration buckets.","tokens_in":13249,"tokens_out":10547,"duration_ms":113244,"significance":"If the construction is made fully specified, the datasets could be a genuinely useful resource: the 1000 FPS acquisition, professional optics, diverse weather/seasonal conditions, multi-task paired sharp/blurry data, and motion-intensity buckets address real limitations of existing benchmarks such as GoPro, Vimeo90K, and Sintel. The reported benchmarking also gives an initial picture of how SOTA methods behave across motion amplitudes. The paper is transparent that the OF labels are pseudo-ground-truth and that the OF results are preliminary. However, the central adaptive-averaging rule is not fully specified, and the OF pseudo-labeling is entangled with the blur-generation procedure, so the claims of consistent blur control and precise multi-task ground truth are not yet established.","major_comments":[{"comment":"The adaptive frame-averaging rule is not specified as a function. The text states a target of 30 pixels of optical flow and gives examples: about 10 px/frame leads to 3 averaged frames, and subpixel motion leads to up to 30 frames. If total blur magnitude is approximately per-frame flow times the number of averaged frames, then a scene with v=0.5 px/frame would require N=60 frames to reach the 30-pixel target, but the stated cap is 30, so the resulting blur would be only about 15 px. Thus the claimed 'consistent blur level' and 'homogeneity' are not guaranteed by the described procedure. The six bins in Fig. 8 and the role of the 'mean and maximum optical flow' mentioned in Sec. 4.3 are also not formally defined. Since frame averaging is the core dataset-construction mechanism, this is a reproducibility and correctness blocker. Please provide the exact mapping, pseudocode, thresholds, an","section":"Sec. 4.3, Fig. 8"},{"comment":"The optical flow benchmark is circular: the pseudo-ground-truth flow labels are generated with DEQFlow [2], which is the same method used in Sec. 4.3 to compute optical flow for selecting the frame-averaging counts. Consequently, the OF evaluation measures agreement with the estimator that defines the blur intensity, and systematic errors in DEQFlow are baked into both the labels and the blur synthesis. The paper should explicitly state this dependence in Sec. 4.3 as well as in Sec. 5.3, and should provide an independent validation subset (e.g., synthetic sequences with known flow, or labels from an independent estimator) to calibrate the pseudo-labels before claiming 'precise ground truth' for optical flow.","section":"Sec. 4.3 / Sec. 5.3"},{"comment":"The reported dataset statistics do not reconcile with the bucket description. Table 3 says VAR-MIORe is organized in buckets from 1 to 249 with 333 input entries per bucket; 249 x 333 = 82,917, not the stated 83,250. It is also unclear whether the seven displayed columns (1, 5, 13, 29, 61, 125, 249) are a subsample of 249 buckets or the only buckets. The 52,218-frame size of MIORe and the per-split counts also cannot be derived from the 333 sequences and the sequence counts in Fig. 8. Please report exact per-bucket counts, the full set of frame-count values, split assignments, and the total number of blurry images for each dataset.","section":"Sec. 4.4, Table 3"},{"comment":"The Total column for VideoFlow is internally inconsistent. Using the split sizes given in Fig. 8 (119 Extreme, 88 Fast, 35 Medium-Fast, 38 Medium, 35 Mild-Medium, 18 Mild), the weighted average of the six EPE values 6.17, 8.44, 9.12, 7.69, 9.57, 8.88 is approximately 7.76, not 19.38. The corresponding F1 total (63.74) does equal the weighted average, and the Total values for the other four methods also behave as weighted averages. Unless a different split or weighting is intended, this appears to be a reporting error in a headline result. Please correct it and state the exact split sizes and weighting used for Total columns.","section":"Table 6"}],"minor_comments":[{"comment":"The PerVFI row cites reference [15], but PerVFI is reference [28] in the bibliography and in Table 4. Please fix the citation.","section":"Table 5"},{"comment":"The figure layout is very hard to read after extraction: the mapping between the six category labels and the two numeric rows (Optical Flow Magnitude and # Sharp Frames) is ambiguous. Provide a clean tabular legend or a separate table with exact ranges and counts.","section":"Fig. 8"},{"comment":"The claim that MIORe and VAR-MIORe are 'the first in the motion deblurring literature to integrate defocus blur as an intrinsic component' is strong and should be supported by a comparison with existing datasets that combine defocus and motion blur, or by a focused literature search.","section":"Sec. 3.3"},{"comment":"The F1 metric for optical flow is not defined. Since the labels are pseudo-ground-truth, please specify the threshold used for F1 and how the EPE is computed (e.g., on which regions, ignoring invalid/occluded pixels).","section":"Sec. 5.3"},{"comment":"The phrase 'sustaining 1000 FPS for both acquisition and final take-away dataset' conflicts with the Table 1 entries 'FPS (Acq. / Final)' of [28,1000] for MIORe and [4,1000] for VAR-MIORe. Clarify whether the final dataset frame rate is the raw 1000 FPS sharp frames or the effective rate of blurry frames.","section":"Sec. 4.1 / Table 1"},{"comment":"Minor wording issue: 'successfully ensuring consistent framing and no motion artifacts' is a run-on; also, 'for several sequences, we use a tripod... successfully ensuring' should be rephrased.","section":"Sec. 4.2"}],"recommendation":"major_revision","confidential_remarks":"This is a potentially valuable dataset paper, but it is not yet self-contained or fully reproducible. The most serious issue is the under-specified adaptive-averaging rule, which is the core mechanism behind both datasets; the OF circularity is secondary but should be framed carefully. The table inconsistencies (Table 6 VideoFlow total, Table 3 bucket counts) should be fixed before publication. If the authors provide exact formulas, per-bucket statistics, split metadata, and an independent OF validation subset, the contribution would likely meet the bar for a dataset benchmark paper. One editorial note: the acknowledgements thank 'reviewers and all members of the program committee'; in a preprint this is unusual and should be removed or rewritten for the review version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should read this one with your dataset hat on. The core idea is solid and fills a real gap: a 1000 FPS multi-task benchmark with explicit control over motion magnitude (VAR-MIORe) is genuinely new, and the adaptive averaging to keep blur intensity roughly constant across slow and fast scenes is a smart move. The benchmarking itself is thorough: five deblurring, five VFI, and five OF methods, with per-split results that show sensible trends. If the data is released and the construction details are pinned down, this could become a useful resource for the restoration community.\n\nThat said, the paper as written is not reproducible. The adaptive rule in Sec 4.3 is described only by examples: high-flow scenes (~10 px/frame) average 3 frames, subpixel scenes average up to 30, targeting 30 px of total flow. If the intended mapping is N = 30/v, then for v < 1 you would need more than 30 frames to hit the 30 px target; the cap at 30 (Fig 8 shows 35 max) means slow scenes get systematically less blur than fast ones. No formula, thresholds, or pseudocode is given, so the claim of 'consistent and controlled motion blur' is unverified.\n\nThe optical flow circularity is real too. The pseudo-labels are computed with DEQFlow, the same method used to estimate flow for the adaptive averaging. So the OF benchmark is essentially evaluating how well methods reconstruct what DEQFlow already 'decided' about the scene, not independent flow. The paper does call them pseudo-labels, but the blur synthesis depends on that same flow, which makes the OF evaluation and the blur generation mutually entangled.\n\nTwo smaller things: the reported statistics don't obviously reconcile (Table 3 says '333 input entries per bucket', but VAR-MIORe total is 83,250 frames; Table 1 lists Size 83,250 for VAR-MIORe and 52,218 for MIORe, yet the split sums to 333 sequences), and no code or data link appears in the arXiv version — only a project page URL.\n\nThese are fixable in revision. The idea is worth a serious referee, and I'd send it out if the authors provide the exact rule, release the data, and address the circularity or at least discuss its implications. As is, the central claim of controlled blur magnitude does not hold quantitatively on inspection.","headline":"A potentially useful multi-task benchmark with a real gap to fill, but the central blur-synthesis rule is under-specified and the OF ground truth is circular; needs a release and a precise formula before it can be trusted.","tokens_in":13670,"tokens_out":2666,"would_cite":false,"duration_ms":32769,"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":"Two 1000-FPS datasets benchmark restoration from sharp to extreme motion","keywords":["motion deblurring","video frame interpolation","optical flow estimation","high-frame-rate dataset","adaptive frame averaging","motion magnitude control","multi-task benchmark"],"falsifier":"Apply the paper's adaptive averaging to synthetic sequences with known true optical flow and measure the resulting blur's flow; a large spread around the claimed 30-pixel target would refute the consistency claim. Separately, recompute optical flow on the sharp frame pairs with a method independent of DEQFlow and compare with the pseudo-labels; systematic deviation would expose the circularity of the optical-flow ground truth.","tokens_in":12831,"feed_emoji":"🎥","tokens_out":4735,"duration_ms":45331,"temperature":0.7,"pith_summary":"The paper introduces MIORe and VAR-MIORe, two multi-task datasets designed to overcome what the authors see as critical gaps in motion-restoration benchmarks: fixed or limited motion magnitudes, lack of controlled blur, and no unified ground truth for deblurring, video frame interpolation, and optical flow. Recorded at 1000 FPS with professional lenses, the datasets generate blurry frames by adaptively averaging a variable number of sharp frames, with the middle sharp frame kept as ground truth. VAR-MIORe extends the range to extreme motion (maximum optical flow of 1932 pixels) and is claimed to be the first benchmark with explicit control over motion amplitude. If these claims hold, researchers gain a single testbed that can reveal where restoration algorithms break as motion intensity grows.","feed_headline":"1000-FPS datasets benchmark restoration from sharp to extreme motion","feed_subtitle":"MIORe and VAR-MIORe give deblurring, frame interpolation, and optical flow one unified testbed with controlled blur and extreme range.","key_machinery":"Adaptive frame averaging driven by the mean and maximum optical flow computed with DEQFlow: fast scenes (about 10 pixels of flow per frame) are averaged over few frames (around 3), while subpixel scenes are averaged over many frames (up to 30), targeting a mid-level optical flow of about 30 pixels per blurry image. This mechanism controls blur magnitude while keeping the left, middle, and right sharp frames as ground truth for VFI, optical flow, and deblurring.","core_discovery":"The central claim is that high-frame-rate acquisition combined with flow-based adaptive frame averaging yields consistent, controlled motion blur while preserving completely sharp flanking and middle frames, enabling the same sequences to serve as ground truth for motion deblurring, video frame interpolation, and optical flow estimation. VAR-MIORe is further claimed to provide the first explicit control over motion amplitude by spanning from static to extreme motion, allowing systematic study of model 'breaking points.' Benchmark results on state-of-the-art methods trained on existing datasets show measurable performance drops across the new datasets, with frequency-domain deblurring methods","pith_inferences":["Because the same DEQFlow method generates the blur (via averaging decisions) and the optical-flow pseudo-labels, the OF evaluation is partly circular; an independent flow estimator would likely change the reported rankings.","The adaptive averaging rule is underspecified — only examples are given, not an exact formula — so reproducibility and the claimed 30-pixel blur target would need a precise mapping to verify.","The 'variable motion amplitude' design could become a standard stress test for restoration models, analogous to rate-distortion sweeps in compression, giving a one-dimensional axis along which to compare robustness.","The observation that frequency-prior methods hallucinate in slow motion implies that benchmarks should deliberately include low-motion regimes, which dominate real consumer footage yet are filtered out by many existing datasets."],"forward_implications":["A single dataset now supplies aligned inputs for deblurring, video frame interpolation, and optical flow, enabling multi-task training and evaluation on identical frames.","VAR-MIORe's continuous motion-magnitude axis lets researchers plot performance versus blur strength and identify the regime where each method breaks.","The reported drop of some deblurring methods on mild motion suggests that training exclusively on fast-motion benchmarks may bias models toward high-frequency priors that misbehave on slow scenes.","Including defocus, exposure, and weather conditions pushes restoration algorithms to handle intertwined optical degradations rather than clean motion only."],"supporting_citations":[{"why":"Supplies the optical flow estimates used both for adaptive frame averaging and for the pseudo-ground-truth optical flow labels.","marker":"[2]"},{"why":"The standard deblurring benchmark whose frame-averaging synthesis is extended; also the training set for the evaluated deblurring models.","marker":"[20]"},{"why":"Prior 1000 FPS extreme-VFI dataset that sets the high-frame-rate acquisition precedent MIORe and VAR-MIORe build on.","marker":"[24]"},{"why":"Reference optical-flow benchmark with ground truth; used for comparison and cited for flow evaluation methodology.","marker":"[3]"},{"why":"Standard VFI benchmark whose moderate-motion curation is contrasted with the new datasets' extreme motion coverage.","marker":"[29]"},{"why":"Real-world blur benchmark used as a comparison for the deblurring claims.","marker":"[21]"},{"why":"Real-world optical-flow benchmark used as a comparison for the OF claims.","marker":"[10]"}],"fun_headline_variants":["New benchmarks expose restoration limits from sharp to extreme motion","1000-FPS benchmarks control blur for deblur, interpolate, flow","MIORe and VAR-MIORe: first explicit motion-amplitude benchmark","Unify deblurring, interpolation, and flow with controlled motion benchmarks"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The rule that chooses how many sharp frames to average into each blurry frame is described only by examples (10 pixels per frame yields 3 frames; subpixel motion yields up to 30) and never as an exact function, so the promised consistent and controlled blur level rests on an ad hoc, underspecified step.","fun_headline_variants_meta":{"raw":{"variants":["New benchmarks expose restoration limits from sharp to extreme motion","1000-FPS benchmarks control blur for deblur, interpolate, flow","MIORe and VAR-MIORe: first explicit motion-amplitude benchmark","Unify deblurring, interpolation, and flow with controlled motion benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001022,"raw_usage":{"total_tokens":4106,"prompt_tokens":662,"completion_tokens":3444,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":406,"completion_tokens_details":{"reasoning_tokens":3364}},"tokens_in":406,"tokens_out":3444,"duration_ms":26222,"temperature":1.0,"reasoning_tokens":3364,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T23:02:15.877141+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply the paper's adaptive averaging to synthetic sequences with known true optical flow and measure the resulting blur's flow; a large spread around the claimed 30-pixel target would refute the consistency claim. Separately, recompute optical flow on the sharp frame pairs with a method independent of DEQFlow and compare with the pseudo-labels; systematic deviation would expose the circularity of the optical-flow ground truth.","supporting_citations":[{"cited_title":"Zico Kolter","cited_arxiv_id":null,"evidence_quote":"Supplies the optical flow estimates used both for adaptive frame averaging and for the pseudo-ground-truth optical flow labels."},{"cited_title":"Deep multi-scale convolutional neural network for dynamic scene deblurring","cited_arxiv_id":null,"evidence_quote":"The standard deblurring benchmark whose frame-averaging synthesis is extended; also the training set for the evaluated deblurring models."},{"cited_title":"XVFI: ex- treme video frame interpolation","cited_arxiv_id":null,"evidence_quote":"Prior 1000 FPS extreme-VFI dataset that sets the high-frame-rate acquisition precedent MIORe and VAR-MIORe build on."},{"cited_title":"Butler, Jonas Wulff, Garrett B","cited_arxiv_id":null,"evidence_quote":"Reference optical-flow benchmark with ground truth; used for comparison and cited for flow evaluation methodology."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Standard VFI benchmark whose moderate-motion curation is contrasted with the new datasets' extreme motion coverage."},{"cited_title":"Real-world blur dataset for learning and benchmarking de- blurring algorithms","cited_arxiv_id":null,"evidence_quote":"Real-world blur benchmark used as a comparison for the deblurring claims."},{"cited_title":"Are we ready for autonomous driving? the KITTI vision benchmark suite","cited_arxiv_id":null,"evidence_quote":"Real-world optical-flow benchmark used as a comparison for the OF claims."}],"review_version":1}