{"id":"cf162ae4-9ad5-413b-85ab-f6b7d45a67cd","arxiv_id":"2507.05594","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A 2D Gaussian video representation with a tri-plane plus polynomial deformation field decodes at 800+ FPS on Bunny and trains in about 2 seconds per frame.","lead":"GSVR represents video as 2D Gaussians that move according to a hybrid deformation field mixing tri-plane grids and polynomial motion. On Bunny it reports 800+ FPS decoding and 35+ PSNR with only about 2 seconds of training per frame.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 800+ FPS decoding claim is plausible per the paper's own measurement, but the reported '2 seconds per frame' encoding time omits RAFT optical-flow preprocessing and the per-GOP re-initialization overhead, so the convergence advantage is not yet fully established.","rationale":"The reader's verdict is CONDITIONAL, and I agree that the main risk is under-specification rather than a clear internal contradiction. The paper's own numbers support an 800+ FPS decoding speed, and the hybrid deformation field has ablations showing that combining tri-plane and polynomial motion improves quality. The central claim that would need to be true is that GSVR trains to 35+ PSNR in 2 seconds per frame and decodes at 800+ FPS with the time-slicing strategy. The most load-bearing unverified component is the encoding-time budget: the dynamic-aware time slicing requires RAFT (Section 3.3), but the implementation details (Section 4.1.3) do not mention RAFT or its runtime, and the optical-flow threshold is not reported. Since training time is the headline contribution, this omission could affect the comparison against NeRV, PNeRV, and HNeRV-Boost. I would keep CONDITIONAL and require a clear accounting of the preprocessing time and threshold. The lack of code and error bars is noted but secondary to the encoding-time concern.","tokens_in":11926,"tokens_out":1612,"duration_ms":16105,"concrete_test":"Recompute Table 2's training-time comparison with RAFT included: run RAFT on Bunny (132x1280x640 frames) and time the optical-flow forward pass plus GOP boundary computation, then add this to the reported per-frame GSVR training time. Also report the adaptive GOP count and the optical-flow threshold; if the total per-frame time including RAFT still matches the baseline setup, the convergence claim stands, otherwise the '2 seconds per frame' claim should be revised.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Table 4 and the abstract compare decoding speed and training time against NeRV, PNeRV, and HNeRV-Boost, where the reported 816.56 FPS and 35+ PSNR are plausible internal measurements. The load-bearing concern is that the head-to-head training-time comparison is incomplete: (1) Section 3.3 requires a pre-trained RAFT optical-flow model to compute per-frame motion magnitude; Section 4.1.3 does not list this forward pass in the training-time budget, yet Table 2 trains models at 2.0 seconds per frame on Bunny. If RAFT is excluded, the actual per-frame encode cost is higher and the claimed '2 seconds per frame' is underestimated. (2) The adaptive GOP slicing also adds multiple independent Gaussian optimizations with per-GOP re-initialization, but the report aggregates training time across all GOPs without stating the GOP count or the per-GOP allocation. (3) Table 5's interpolation claim compares only with HNeRV-Boost and notes that for HNeRV-Boost test frames would be needed during evaluation; even with that caveat, Section 4.5.1 does not specify how odd/even frame splitting interacts with the optical-flow-based GOP boundaries, so the interpolation protocol may not be reproducible. (4) The optical-flow threshold that segments GOPs is never reported, which is noted by the reader. Because the dynamic-aware slicing is load-bearing for both reconstruction quality and compression performance, its unreported threshold and excluded RAFT time leave the headline encoding-time claim and the comparison with baselines under-specified. These are issues of missing disclosure and potential underestimation, not internal inconsistency.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GSVR, a 2D-Gaussian-based video representation with a hybrid deformation field combining tri-plane grids and polynomial motion, a dynamic-aware time-slicing strategy based on optical flow to segment the video into GOPs, and a compression pipeline using quantization-aware fine-tuning and image codecs. The central claims are that GSVR achieves 800+ FPS decoding, 35+ PSNR on Bunny, and roughly 2 seconds per frame training time, with faster convergence than NeRV-style baselines. Experiments on Bunny and UVG report PSNR, FPS, interpolation, and compression results against NeRV, PNeRV, and HNeRV-Boost, plus an ablation of GOP length. The method is clearly described and the reported internal measurements are plausible, but the training-time accounting omits the RAFT optical-flow preprocessing required by the time-slicing strategy, and several reproducibility-critical details (GOP threshold, parameter accounting across GOPs, interpolation protocol) are missing.","tokens_in":12371,"tokens_out":5662,"duration_ms":65669,"significance":"If the training-time and FPS claims hold after a full accounting of preprocessing and model-structure costs, GSVR would be a substantial advance over convolution-based neural video representations: it removes the MLP/convolution decoder bottleneck, reaches two orders of magnitude higher decode throughput, and converges to high PSNR far earlier than NeRV-style baselines. The paper also contributes a hybrid motion model that explicitly separates camera-like and object-like motion through a per-Gaussian dynamic indicator, and a compression scheme with quantization-aware fine-tuning. The strengths are the clarity of the core representation, the direct FPS measurements on two datasets, and the inclusion of a GOP-length ablation. However, the headline '2 seconds per frame' and the faster-convergence comparison are not yet fully supported because the RAFT preprocessing cost and the number/parameters of GOPs are not reported; the comparison also omits the closest concurrent 2D-Gaussian video methods. These are fixable with additional reporting and targeted experiments, not fundamental flaws in the representation.","major_comments":[{"comment":"The headline training-time comparison is not end-to-end. Section 3.3 requires a pre-trained RAFT optical-flow forward pass per frame to compute the motion magnitude D, but Section 4.1.3 reports no RAFT details and Section 4.1.2 defines training time as total training seconds divided by total frame numbers without mentioning preprocessing. The adaptive GOP threshold is also never reported. Because Tables 2 and 3 plot PSNR against this training-time axis, the claimed '2 seconds per frame' and the faster-convergence comparison are currently under-specified. Please report the end-to-end encoding time including RAFT and per-GOP re-initialization, or explicitly state that preprocessing is excluded and quantify its cost, and give the threshold and resulting GOP boundaries for each sequence.","section":"§3.3, §4.1.2, §4.1.3, Tables 2–3"},{"comment":"The parameter accounting across GOPs is unclear. Section 3.3 states that each GOP is represented by independent 2D Gaussians and a deformation field, while Section 4.1.2 says all models are 3M parameters. If each GOP is independent, the total parameter count would scale with the number of GOPs unless the budget is shared; if the 3M budget is shared, the per-GOP model size varies and must be specified. This ambiguity affects the training-time comparison in Tables 2–3 and the bits-per-pixel compression numbers in Table 7, both of which are load-bearing for the paper's central claims.","section":"§3.3, §4.1.2, Table 7"},{"comment":"The interpolation protocol is not reproducible as written. The paper uses odd frames for training and even frames for testing, but Section 3.3's GOP segmentation is based on optical flow computed on the video; the text does not state whether flow was computed on the full video, whether GOP boundaries were aligned with the train/test split, or how many frames fell in each set. In addition, the note that HNeRV-Boost requires test frames to form frame embeddings conflicts with the test-set PSNR reported in Table 5. Please specify the exact protocol for both methods and, if test frames were used for HNeRV-Boost, state this explicitly and discuss the implication for the comparison.","section":"§4.5.1, Table 5"},{"comment":"The main FPS and quality tables omit the closest concurrent methods. Section 2.2 names D2GV and GaussianVideo as deformable 2D-Gaussian video representations, and D2GV is reported to reach 400 FPS; nevertheless, Tables 1–4 compare only NeRV, PNeRV, and HNeRV-Boost. The abstract's 10x decoding-speed claim and the method's positioning therefore need at least a quantitative comparison or a clear explanation of why D2GV/GaussianVideo cannot be compared. At minimum, include their published FPS, PSNR, and training-time numbers in the discussion.","section":"§2.2, §4.2–4.3, Tables 1–4"}],"minor_comments":[{"comment":"The notation in Eq. (4) uses a product over three plane interpolations, but the text later describes a fusion and Eq. (10) uses a weighted sum; please clarify whether the plane features are multiplied, summed, or concatenated before producing the attribute offsets.","section":"§3.2, Eq. (4)"},{"comment":"The per-Gaussian dynamic indicator alpha is said to be learnable, but its range and initialization are not stated; if it is meant to be a convex combination weight, please constrain or regularize alpha to [0,1] and describe the initialization.","section":"§3.2, Eq. (10)"},{"comment":"The 'Adaptive' row in Table 6 reports the best mean PSNR, but the paper does not give the adaptive threshold or the resulting GOP lengths for any UVG sequence; please report these values so the adaptive slicing can be reproduced.","section":"§4.4.2, Table 6"},{"comment":"Table 7 does not specify which image codec (PNG or JPEG-XL) and settings were used, nor whether the reported bits per pixel include both the Gaussian attribute grids and the tri-plane grids; please clarify, and add an ablation comparing quantization-aware fine-tuning with post-hoc quantization.","section":"§3.4, Table 7"},{"comment":"There are several typographical errors and inconsistent hyphenations, including 'utlize' in §3.3, 'polymonical' in Figure 3, and 'indictor' in §3.2; a careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The core representation is promising and the reported decode speed is plausible, but the training-time claim and cross-method comparisons need substantial additional reporting before publication. The missing RAFT preprocessing time, GOP threshold, and parameter accounting are fixable in revision; the authors should also add a direct or tabulated comparison to D2GV and GaussianVideo, since those are the closest prior works on 2D-Gaussian video representation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two quick takes. The 800+ FPS claim is believable — the architecture is a clean MLP-free 2D Gaussian splatting with a hybrid deformation field, and the measured 816 FPS on Bunny is consistent with how fast GaussianImage already is. The second take is that the paper's training-time comparison is under-specified: the preprocessing RAFT optical flow from Section 3.3 is not included in the per-frame training budget, and the GOP-slicing threshold is never reported. Without those, reproducing the '2 seconds per frame' headline is not possible.\n\nWhat's actually new: the hybrid deformation field — tri-plane for background/camera motion, degree-2 polynomial for dynamic objects, and a learnable per-Gaussian indicator to mix them — plus the optical-flow-driven adaptive GOP slicing. The ablation in Figure 6 and Table 6 shows these choices matter, and the gains over NeRV-style baselines in both PSNR and convergence speed are large. The compression numbers (3.1 MB for Bunny's 3M parameters) are also reasonable.\n\nThe weak spots are mostly missing disclosure. The RAFT forward pass could easily add 0.1–0.3 s/frame on a 720p frame, which is material if you're claiming 2.0 s/frame. The GOP threshold is a free parameter that affects both quality and compression, so it needs to be stated. The interpolation test uses only HNeRV-Boost and doesn't specify how the odd/even split interacts with GOP boundaries; the claim of 'SOTA interpolation' is overreaching. And the absence of D2GV, the closest concurrent work, from the comparisons is a real gap for a paper whose headline is speed.\n\nNone of this is fatal. The method is coherent, the internal measurements are consistent, and the central claim about decoding speed is plausible. The paper is worth a serious referee, but it needs revision for transparency: report the threshold, include RAFT in the timing budget or state it's excluded, add D2GV, and soften the interpolation claim.","headline":"A credible, fast-decoding video representation with clear design choices, but the training-time claim omits optical-flow preprocessing and the adaptive GOP threshold is undisclosed, so it needs transparency fixes before publication.","tokens_in":12815,"tokens_out":2843,"would_cite":true,"duration_ms":29416,"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 a 2D Gaussian-based video representation that decodes at 800+ FPS on Bunny and reaches 35+ PSNR in about 2 seconds of training per frame.","keywords":["2D Gaussian splatting","implicit neural representation","video representation","deformation field","video compression","real-time decoding","optical flow","quantization-aware fine-tuning"],"falsifier":"Run the full encoding pipeline on Bunny with a timer that includes RAFT optical-flow computation and the adaptive GOP segmentation, and compare total encode time against the reported 2 seconds per frame; also check whether 816.56 FPS still holds when tri-plane bilinear interpolation and Gaussian rasterization are included in the timed forward pass.","tokens_in":11710,"feed_emoji":"🎥","tokens_out":4448,"duration_ms":44835,"temperature":0.7,"pith_summary":"This paper aims to show that implicit neural video representations do not have to be slow: it proposes GSVR, a video representation built from 2D Gaussians instead of convolutional NeRV blocks. The central claim is that on the Bunny sequence GSVR decodes at over 800 FPS and reaches 35+ PSNR with only about 2 seconds of training per frame, roughly 10x faster decoding than NeRV-based baselines with much faster convergence. A hybrid deformation field splits motion into tri-plane motion for camera and background and polynomial motion for fast-moving objects, so the representation separates static and dynamic content without any MLP. Dynamic-aware time slicing uses optical flow to cut the video into groups of pictures, and quantization-aware fine-tuning plus image codecs compress the Gaussians. The authors also report competitive video interpolation and better compression than NeRV.","feed_headline":"2D Gaussian video representation decodes at 800+ FPS","feed_subtitle":"Hybrid deformation field and optical-flow time slicing reach 35+ PSNR in 2 seconds of training per frame","key_machinery":"The load-bearing object is the canonical 2D Gaussian, an ellipse with position, scale, rotation, and color, rasterized by weighted blending in screen space. Around it sits the hybrid deformation field, which is fully explicit: tri-plane grids return attribute offsets by bilinear interpolation of $(x,y)$, $(x,t)$, and $(y,t)$ planes, and a learnable quadratic polynomial returns a position offset for fast motion; a per-Gaussian dynamic indicator $\\alpha$ fuses the two. Because there is no MLP anywhere in the deformation path, decoding reduces to grid lookups, polynomial evaluation, and Gaussian blending. The dynamic-aware time slicing strategy supplies the video-level structure, using the mean absolute optical flow per frame to accumulate a motion score and cut GOP boundaries when it crosses a threshold.","core_discovery":"On its own terms, the paper's discovery is that 2D Gaussian splatting, previously used for images, can serve as the backbone of a video representation that is simultaneously fast to train and fast to decode. Each frame is produced by rasterizing a canonical set of 2D Gaussians deformed through time by a hybrid deformation field: tri-plane grids capture camera motion and slowly changing regions, a quadratic polynomial term captures high-dynamic objects, and a learnable dynamic indicator per Gaussian blends the two. Removing the MLP decoder that most deformation fields use keeps inference cheap, and dividing the video into adaptive GOPs based on optical-flow magnitude prevents long sequences from degrading. The reported numbers on Bunny are 816.56 FPS decoding and 35.48 PSNR at 2.0 training seconds per frame, versus 84.55 FPS for NeRV; on UVG the method averages 538.49 FPS and 30.22 PSNR.","pith_inferences":["The reported 2-second-per-frame training time likely omits the cost of computing RAFT optical flow for dynamic-aware time slicing, since the paper never mentions its runtime; the true end-to-end encoding cost is probably higher.","The optical-flow threshold that decides GOP boundaries is never reported, so the adaptive slicing cannot be reproduced exactly from the text; reporting it would make the strategy deterministic and comparable.","One testable extension is to replace the quadratic polynomial with higher-order or piecewise motion for sequences with repeated fast motion, and to check whether the dynamic indicator still cleanly separates foreground from background.","The hybrid deformation idea transfers naturally to dynamic-scene representations that currently rely on MLP deformation fields, potentially giving them the same decoding-speed boost."],"forward_implications":["Decoding exceeds 60 FPS at 1080p-class resolutions on the tested GPU, so real-time playback with random frame access is feasible for neural video representations.","Training time near 2 seconds per frame makes the representation practical for episodic or streaming content rather than offline-only processing.","The same canonical Gaussians plus explicit deformation support video interpolation with temporal continuity, since motion is modeled rather than baked into frame embeddings.","Quantization plus image-codec compression yields compact bitrates, making the representation a candidate for neural codecs that reuse standard still-image encoders.","Because the representation separates dynamic and static elements, downstream editing tasks could target the moving object by adjusting its polynomial coefficients and dynamic indicator."],"supporting_citations":[{"why":"Supplies the canonical 2D Gaussian representation and rasterization pipeline that GSVR adopts as its base.","marker":"[37]"},{"why":"Defines NeRV, the convolution-based video representation that GSVR replaces and the main speed baseline.","marker":"[4]"},{"why":"Provides RAFT optical flow, used to estimate per-frame motion for the dynamic-aware time slicing strategy.","marker":"[27]"},{"why":"Inspires the sliding-window GOP idea for dynamic 3D Gaussian splatting, which GSVR adapts with optical-flow-based boundaries.","marker":"[23]"},{"why":"Example of a grid-based deformation field for dynamic scenes, whose MLP decoder GSVR removes for speed.","marker":"[31]"},{"why":"Concurrent Gaussian-based video representation used as a comparison point for compression performance.","marker":"[11]"},{"why":"HNeRV-Boost, a stronger NeRV baseline used in reconstruction and interpolation comparisons.","marker":"[38]"},{"why":"HNeRV, a frame-embedding NeRV variant used as a baseline for quality and interpolation.","marker":"[3]"},{"why":"Provides the idea of mapping unordered Gaussians to 2D grids so image codecs can compress them.","marker":"[19]"}],"fun_headline_variants":["2D Gaussians decode video at 800+ FPS","Video trains in 2s/frame, decodes at 800+ FPS","2D Gaussian video reaches 35+ PSNR at 800+ FPS","Hybrid deformation field enables 800+ FPS video decode","Hybrid deformation field powers 800+ FPS video representation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported 2-second-per-frame training time appears to exclude the cost of computing RAFT optical flow used by the dynamic-aware time slicing, and the optical-flow threshold that sets GOP boundaries is never reported.","fun_headline_variants_meta":{"raw":{"variants":["2D Gaussians decode video at 800+ FPS","Video trains in 2s/frame, decodes at 800+ FPS","2D Gaussian video reaches 35+ PSNR at 800+ FPS","Hybrid deformation field enables 800+ FPS video decode","Hybrid deformation field powers 800+ FPS video representation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000974,"raw_usage":{"total_tokens":4187,"prompt_tokens":1039,"completion_tokens":3148,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":655,"completion_tokens_details":{"reasoning_tokens":3052}},"tokens_in":655,"tokens_out":3148,"duration_ms":21958,"temperature":1.0,"reasoning_tokens":3052,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:22:17.176525+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the full encoding pipeline on Bunny with a timer that includes RAFT optical-flow computation and the adaptive GOP segmentation, and compare total encode time against the reported 2 seconds per frame; also check whether 816.56 FPS still holds when tri-plane bilinear interpolation and Gaussian rasterization are included in the timed forward pass.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the canonical 2D Gaussian representation and rasterization pipeline that GSVR adopts as its base."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines NeRV, the convolution-based video representation that GSVR replaces and the main speed baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides RAFT optical flow, used to estimate per-frame motion for the dynamic-aware time slicing strategy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Inspires the sliding-window GOP idea for dynamic 3D Gaussian splatting, which GSVR adapts with optical-flow-based boundaries."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Example of a grid-based deformation field for dynamic scenes, whose MLP decoder GSVR removes for speed."},{"cited_title":"GaussianVideo: Efficient Video Representation and Compression by Gaussian Splatting","cited_arxiv_id":"2503.04333","evidence_quote":"Concurrent Gaussian-based video representation used as a comparison point for compression performance."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"HNeRV-Boost, a stronger NeRV baseline used in reconstruction and interpolation comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"HNeRV, a frame-embedding NeRV variant used as a baseline for quality and interpolation."}],"review_version":1}