{"id":"9b384b9f-fa19-441a-af72-4008b7bc1f52","arxiv_id":"2504.14535","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FlowLoss adds a noise-gated optical-flow matching loss to video diffusion training, yielding earlier motion stability but mixed final results and higher training cost.","lead":"This paper proposes FlowLoss, a training loss for video diffusion models that compares optical flow fields of generated and real videos, and only applies it when noise is low. It reports faster early training convergence and more stable motion, but no consistent improvement at convergence and longer training time.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Early stabilization may be static-bias, not learned motion: the flow loss penalizes deviation from near-zero ground-truth flow on a fixed-view, mostly-static dataset, and the step-count comparison ignores large per-step compute overhead.","rationale":"I read the paper as a modest proposal: use differentiable flow matching with noise-aware gating to improve early temporal coherence. The method is clearly described and the limitations are candid. However, the central claim rests on early-training evidence that is not yet disentangled from the dataset's static bias. On fixed-view Bridge v2, ground-truth flow is near zero; direct flow matching can reward static outputs, so \"stability\" at step 100 may be freezing rather than motion learning. The authors themselves note the flow extractor may be noisy in occluded or low-texture regions, which would only strengthen this bias. Also, since FlowLoss roughly doubles per-step cost for larger ψ, step-count comparisons overstate practical acceleration. These are not accusations of dishonesty; they are missing controls. The reader's weakest assumption (flow-estimator gradient reliability) is related but not identical; my concern is about interpreting the observed early curves even if the gradients are well-behaved. I therefore partially agree with the reader. The appropriate verdict remains CONDITIONAL: the paper should be accepted only if early-stage metrics with error bars, motion-magnitude measurements, and compute-controlled comparisons confirm the claim.","tokens_in":7145,"tokens_out":7131,"duration_ms":70625,"concrete_test":"Train baseline and FlowLoss (e.g., ψ=0.125) with 3 seeds. At steps 100, 500, 2,000, and 5,000, report (i) FVD with confidence intervals; (ii) mean optical-flow magnitude of generated videos measured with a fixed flow estimator, alongside ground-truth flow magnitude; (iii) wall-clock time and FLOPs per 100 steps. If FlowLoss's early generated motion magnitude is far below ground truth while baseline shows jitter but comparable magnitude, the \"motion prior\" claim is a static-bias artifact. If time-to-reach the baseline's step-100 FVD is not shorter for FlowLoss, the acceleration claim fails in practical terms.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that FlowLoss \"improves motion stability and accelerates convergence in early training stages.\" The supporting evidence is Fig. 4 (training curves, no error bars or seeds) and Fig. 5 (selected samples at step 100). On Bridge v2 the camera is fixed and \"most pixels are static\" (Sec. 4), so ground-truth flow is near zero. Lflow directly penalizes any deviation from this flow, so an early model that outputs a mostly static video receives a low flow loss. The observed step-100 stabilization may therefore be a static-bias artifact: the model is pushed toward \"no motion\" rather than toward the correct robot-arm motion prior. This is reinforced by the occlusion/low-texture caveats in Sec. 4. A second, compounding problem is that Table 1 shows FlowLoss adds 10-82% wall-clock time per training run; comparing by step count means \"accelerates convergence\" can reflect extra compute per step, not better sample efficiency. Because Lflow is also the evaluation metric in Fig. 4(f), faster Lflow minimization is expected and is not independent confirmation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FlowLoss, an auxiliary training objective for video diffusion models that extracts optical flow from both the denoised prediction and the ground-truth video with a pretrained estimator (DOT), compares the two flow fields directly, and gates the resulting loss by noise level through a hard threshold ψ and a global scale s. The full objective is L = L_recon + w_ψ(σ)·L_flow, where L_flow is only active at low noise levels. Experiments on Bridge v2 with a Stable Video Diffusion 2.1 UNet compare the reconstruction-only baseline with several FlowLoss variants. The authors report early-training stabilization visible in validation curves (Fig. 4) and sample generations at step 100 (Fig. 5), while candidly acknowledging that final quantitative results are mixed and that computation time increases substantially.","tokens_in":7351,"tokens_out":3556,"duration_ms":34524,"significance":"If the central claim held, the paper would make a modest but useful contribution: a direct-flow-matching loss with noise-aware gating is a reasonable alternative to warping-based flow supervision for video diffusion training, and the explicit analysis of low-noise versus high-noise supervision is valuable. The authors are unusually transparent about the limitations of their evidence, including the absence of consistent final-stage gains and the computational overhead. However, the early-training benefit is supported mainly by qualitative samples and by a validation metric that is essentially the training loss itself, on a single dataset and without repeated seeds. The idea is worth pursuing, but the manuscript currently does not establish the central claim with independent, quantitative evidence.","major_comments":[{"comment":"Fig. 4(f) plots Lflow/λ(σ) as a validation metric, but this is essentially the same flow-matching quantity that the proposed training objective minimizes (with gating and scaling). A faster decrease of this quantity under FlowLoss is therefore expected by construction and does not independently confirm that the model has learned a better motion prior. The paper reports 'unweighted Lflow' as its temporal metric in §4 and Fig. 4(f) labels Lflow/λ(σ); either way, this metric is not independent of the training signal. To support the acceleration claim, the authors should report independent temporal consistency metrics (e.g., FVD on dynamic regions, temporal warping error, or human evaluation) with confidence intervals over multiple seeds.","section":"§4, Fig. 4(f)"},{"comment":"The early stabilization reported in Fig. 5 may be a static-bias artifact rather than evidence of learning a correct motion prior. On Bridge v2 the camera is fixed and, as the authors note in §4, 'most pixels are static.' With ground-truth flow fy near zero, the flow loss in Eq. (1) approximately reduces to s·E||f_yhat||², which directly penalizes any generated motion. A model that outputs a mostly static video will therefore receive a low Lflow and may appear 'stable' at step 100 without having learned the robot-arm motion. The authors should test FlowLoss on a dataset with substantial object or camera motion, or add a control experiment that penalizes temporal differences without using flow, and report metrics restricted to dynamic regions.","section":"§3, Eq. (1); §4, static-bias concern"},{"comment":"All convergence comparisons are by training step, but Table 1 shows that the FlowLoss runs add 10–82% wall-clock time (e.g., 14h29m for ψ=0.250 versus 7h56m for the baseline). Comparing by step count therefore conflates 'better sample efficiency' with 'more compute per step.' The paper's practical claim of accelerated convergence should be evaluated by early-stage quality as a function of wall-clock time or total compute, not just step number. The authors should also provide exact early-stage numbers (e.g., FVD, LPIPS at step 100) with multiple seeds; the plotted curves in Fig. 4 appear to be single runs with no error bars.","section":"§4, Table 1 and Fig. 4"},{"comment":"The method depends on several hand-chosen constants: the global scaling factor s = 10⁻⁶, the occlusion fallback weight 0.3, and the gating threshold ψ. The reported experiments vary ψ, but provide no sensitivity analysis for s or the occlusion weight, and the ablation in §5 uses a different 'Weighted Average' formulation rather than the default additive objective. Since these constants directly control the magnitude and active region of the flow loss, at least a small sensitivity study (and ideally repeated-seed training) is needed to show that the early-training benefit is robust rather than tuned to the particular choices.","section":"§3, Eqs. (1)–(2); §5"}],"minor_comments":[{"comment":"The definition of the occlusion mask says 'o(αt(i,j)) = 1 for visible regions and 0.3 otherwise'; the occlusion fallback value 0.3 is not justified and the notation would benefit from an explanation of why a nonzero weight is chosen for occluded pixels.","section":"§3, Eq. (1)"},{"comment":"The four panels in Fig. 2 are very small and the curves are hard to read; in particular, panel (c) and (d) would benefit from larger fonts and clearer legends, since the comparison of weighting strategies is central to the method.","section":"Fig. 2"},{"comment":"The numeric values in Table 1 lack spaces after some commas (e.g., '0.811828.8986' should be parsed as SSIM 0.8118 and PSNR 28.8986), making the table difficult to read. Also, the bold highlighting should be explained in the caption.","section":"Table 1"},{"comment":"There is a grammatical issue in the second paragraph of §2: 'Dense optical flow captures pixel-level motion dynamics, it serves as a valuable signal' should be split into two sentences. Additionally, references [9] and [13] appear to describe the same work (optical-flow guided prompt optimization) with different IDs; the authors should verify and merge duplicates.","section":"§2, Related Work"},{"comment":"The description of 'unweighted Lflow' as a metric is ambiguous because Eq. (1) defines Lflow with the λ(σ) weighting and global scale s; the authors should state exactly what is computed for evaluation, including whether the scale and occlusion mask are applied during evaluation.","section":"§4, Evaluation Metrics"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is honest and clearly written, but the evidence is thinner than the abstract claims: one dataset, one seed, no error bars, and the main early-training metric is partly the training objective itself. I would not reject the submission, but the central claim needs to be re-established with independent metrics, wall-clock comparisons, and ideally a dataset with non-trivial motion. I also suggest the authors release code and check the duplicate references before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: FlowLoss is a real alternative to warping-based flow supervision — direct flow-field matching with a noise-gated schedule — but its headline claim of faster early convergence rests on shaky evidence. On Bridge v2, a fixed-camera dataset where most pixels are static, the flow loss penalizes any deviation from near-zero ground-truth flow, so a model that outputs mostly static video receives a low loss. That makes the early Lflow curve fall faster by construction, and the FVD samples at step 100 may just show the model learning to freeze rather than to move correctly. The authors even note the dataset is 'mostly static' and that the model 'adapt[s] to its static layout,' which is the alternative explanation the paper does not rule out.\n\nWhat is genuinely new: the formulation itself. Prior flow-supervision works (FlowVid, LFDM) use warping objectives or treat flow as conditioning; here the flow fields from generated and ground-truth videos are compared directly, with a hard gate that only activates Lflow below a noise threshold. The wss-vs-wls ablation is a clean test of the low-noise-useful intuition and supports it. The paper is also refreshingly candid: they state outright that final-stage metrics show no consistent advantage and that training time increases by 10–82%.\n\nThe soft spots are substantial but not fatal. The early-training benefit is supported by curve shapes from a single seed, no error bars, and one dataset. Lflow is both the training objective and an evaluation metric in Fig 4f, so faster minimization is expected and is not independent evidence. The step-count comparison ignores per-step compute overhead; 82% extra wall-clock time makes 'accelerates convergence' ambiguous. No code or prior flow-supervision baselines are included. These are addressable, and the authors' own limitation discussion keeps the paper credible.\n\nNet: this is a modest, honest contribution. It would benefit from a follow-up that distinguishes 'learns motion' from 'learns stillness' — for instance, a dynamic-camera dataset or an evaluation that separates static background from moving object pixels. If you are building video diffusion for robotics, the idea is worth a read; the evidence as it stands should not drive design choices yet.\n\nI would send this to peer review with a referee note asking for the static-bias analysis and per-step efficiency comparison.\n\nBest.","headline":"A genuine but modest idea whose main claimed benefit is confounded by the static-bias problem and extra compute, but the paper is honest and deserves a critical referee.","tokens_in":7865,"tokens_out":2076,"would_cite":false,"duration_ms":18911,"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":"Direct optical-flow comparison, gated by noise, steadies early video-diffusion training.","keywords":["FlowLoss","video diffusion models","optical flow supervision","noise-aware weighting","motion stability","training convergence","robotic manipulation videos","flow loss gating"],"falsifier":"Train the same model on a robotic dataset with large textureless or heavily occluded regions while keeping FlowLoss; if the early-training flattening of FVD and flow-error curves disappears or flips, the claim that flow supervision itself drives early stabilization fails. Alternatively, replace DOT flow fields with randomized flow fields of the same magnitude; if early training still stabilizes, the effect comes from the loss shape, not from real motion information.","tokens_in":6942,"feed_emoji":"🎬","tokens_out":5290,"duration_ms":45580,"temperature":0.7,"pith_summary":"Video diffusion models generate realistic frames but often produce jittery or physically inconsistent motion. FlowLoss attacks this by adding a training term that directly compares optical flow extracted from generated videos with flow from ground-truth videos, instead of using warping-based objectives. Because flow estimates are unreliable on heavily noised inputs, the loss is gated by the diffusion noise level so it acts mainly in low-noise regimes. On robotic manipulation videos, this scheme makes training stabilize earlier and reach a reasonable motion prior faster, although final-generation metrics do not consistently beat the reconstruction-only baseline.","feed_headline":"Flow loss steadies early video-diffusion training","feed_subtitle":"Comparing generated and real optical flow, gated by noise, speeds convergence to a coherent motion prior.","key_machinery":"The method couples the EDM-style diffusion objective $L_{\\mathrm{recon}} = \\mathbb{E}[\\lambda(\\sigma)\\lVert D_\\theta(y+n;\\sigma)-y\\rVert^2]$ with a flow loss $L_{\\mathrm{flow}} = s\\cdot\\mathbb{E}[\\lambda(\\sigma)\\sum_t o(\\alpha_y^t)\\lVert f_y^t - f_{\\hat{y}}^t\\rVert_2^2]$, where $f$ is a pretrained dense flow extractor (DOT, Dense Optical Tracking), $o$ is an occlusion mask, $s=10^{-6}$, and $\\lambda(\\sigma)=(\\sigma^2+1)/\\sigma^2$. A hard gating function $w_\\psi(\\sigma)=1/(\\sigma^2+1)$ for $\\sigma<\\psi$ and $0$ otherwise decides when the flow term participates, so supervision concentrates on clean inputs where flow is reliable.","core_discovery":"The paper's central claim is that direct, noise-conditioned optical-flow comparison can serve as an effective auxiliary objective for video diffusion training. Concretely, with the flow loss active, validation curves for FVD (a video-quality metric) and flow error flatten sooner, and samples at step 100 show less spatial drift and jitter than the reconstruction-only baseline. The authors attribute the early benefit to the flow signal steering the model toward a coherent motion prior in the low-noise stages where denoising is most sensitive, while acknowledging that hard gating and noisy flow estimates in occluded or texture-free regions limit the final gains.","pith_inferences":["Because the benefit is concentrated at the start of training, a natural extension is to schedule flow loss with a decaying weight or disable it after a motion prior is established, trading the compute overhead for the early gain.","In static-camera robotic scenes, most pixels barely move; direct flow comparison may mostly reinforce static layout, which could explain why final metrics do not improve. Applying FlowLoss on dynamic scenes or with a flow estimator robust to textureless regions is a testable follow-up.","The same noise-gated supervision idea could transfer to other motion-related auxiliary signals, such as point trajectories or depth changes, for video diffusion training."],"forward_implications":["Early in training, flow supervision flattens FVD and flow-error validation curves sooner, indicating a faster-formed motion prior.","Gating flow loss to low-noise steps avoids harmful gradients; applying it at high-noise steps produces flickering, unstable outputs.","The benefits are largely confined to early training; final metrics are comparable or mixed, and training time grows substantially with flow loss.","A weighted-average variant that emphasizes flow at small $\\sigma$ matches the default method's early stability, reinforcing the noise-aware scheduling intuition."],"supporting_citations":[{"why":"Provides the updated image-to-video implementation and FiLM conditioning setup the authors build on.","marker":"[24]"},{"why":"Supplies the EDM diffusion training objective and the $\\sigma$-dependent weighting $\\lambda(\\sigma)$ that FlowLoss reuses.","marker":"[30]"},{"why":"Supplies the pretrained dense flow estimator (DOT) that extracts flow fields from both predictions and ground truth.","marker":"[31]"},{"why":"Provides the Stable Video Diffusion 2.1 UNet backbone used for the video diffusion model.","marker":"[32]"},{"why":"Provides the Bridge Dataset v2 robotic manipulation videos used for training and evaluation.","marker":"[33]"}],"fun_headline_variants":["Noise-gated flow loss speeds video diffusion convergence","Direct flow loss stabilizes early video diffusion training","Flow loss with noise weighting improves motion coherence","Flow loss reduces drift in video diffusion early training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the pretrained flow estimator, applied to denoised and ground-truth videos at low noise levels, produces gradients that point toward coherent motion rather than artifacts, so that the benefit vanishes if flow estimates are noisy in textureless or occluded scenes.","fun_headline_variants_meta":{"raw":{"variants":["Noise-gated flow loss speeds video diffusion convergence","Direct flow loss stabilizes early video diffusion training","Flow loss with noise weighting improves motion coherence","Flow loss reduces drift in video diffusion early training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000245,"raw_usage":{"total_tokens":1454,"prompt_tokens":783,"completion_tokens":671,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":399,"completion_tokens_details":{"reasoning_tokens":612}},"tokens_in":399,"tokens_out":671,"duration_ms":6276,"temperature":1.0,"reasoning_tokens":612,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:45:44.114938+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same model on a robotic dataset with large textureless or heavily occluded regions while keeping FlowLoss; if the early-training flattening of FVD and flow-error curves disappears or flips, the claim that flow supervision itself drives early stabilization fails. Alternatively, replace DOT flow fields with randomized flow fields of the same magnitude; if early training still stabilizes, the effect comes from the loss shape, not from real motion information.","supporting_citations":[{"cited_title":"Dense optical tracking: Connecting the dots","cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained dense flow estimator (DOT) that extracts flow fields from both predictions and ground truth."}],"review_version":1}