{"id":"120f7204-e637-427a-b63b-c8d26c83a8ee","arxiv_id":"2608.10439","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Stream Forcing constructs a curriculum training trajectory in a Logit-normal parameterized sampling space to reconcile training coverage with inference consistency for streaming video diffusion, reporting FVD improvements.","lead":"This paper introduces a training schedule for streaming video diffusion models that gradually shifts from independent per-frame noise sampling to inference-aligned progressive sampling. The method reports large FVD gains on UCF-101, but the headline improvement compares against a baseline trained with test-set leakage.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 36.6% FVD headline compares to FAR*, which Table 1 marks as trained on train+test; against train-only AR-Diffusion† the gain is ~2.7%, so the central claim rests on an inconsistent baseline.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the headline 36.6% improvement is computed against FAR*, which Table 1 marks as trained on both train and test splits, while Stream Forcing is train-only. The paper's own table footnote makes this mismatch explicit, so the claimed 'same evaluation setting' in Section 4.1.1 is internally inconsistent. With the train-only SOTA AR-Diffusion† at 181.9, the same-setting gain is only about 2.7%, which is a materially different claim. I did not find a separate mathematical flaw: the Logit-Normal reformulation is coherent, the mode equation derivation in A.4 is self-contained, and the ablations move in the expected direction. The A.6.1 limitation statement about small/medium-scale experiments reinforces caution but is secondary. Because the flaw is in evaluation reporting rather than the core training construction, the method may still be viable, and the reader's CONDITIONAL verdict remains appropriate; no verdict change is needed. The long-horizon 27.9% claim is less affected because its Table 2 comparison is against train-only Diffusion Forcing.","tokens_in":15856,"tokens_out":3830,"duration_ms":32847,"concrete_test":"Re-run FAR* [13] under the identical train-only protocol used for Stream Forcing, keeping the same backbone, resolution, metric, and evaluation split as Table 1; if the train-only FAR* FVD is much larger than 279.0, recompute the claimed improvement against the train-only SOTA (AR-Diffusion†, 181.9). Also report per-seed or bootstrap confidence intervals for Ours, AR-Diffusion†, and FAR* so the ~2.7% gap can be distinguished from noise. If a train-only FAR* run is unavailable, the paper should re-report its headline comparisons using only train-only baselines.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline 36.6% FVD improvement is (279.0 - 177.0) / 279.0, computed against FAR*, but Table 1 explicitly marks FAR* with '*' (trained on both train and test splits), while Stream Forcing is trained on train only. Section 4.1.1 claims the comparison is 'under the same evaluation setting,' which the table contradicts. Against the best train-only baseline, AR-Diffusion† at 181.9, the improvement is (181.9 - 177.0) / 181.9 ≈ 2.7%, not 36.6%. Because the abstract, introduction, and experiments all lead with the 36.6% number, the central empirical claim depends on a baseline with test-set leakage. The method's internal ablations (Tabs 4-7) are consistent and the mode equation proof in A.4 is sound, so this is not a mathematical defect; it is an evaluation comparability defect. A.6.1 also concedes that experiments are small- to medium-scale, which further limits the strength of the headline. The long-horizon 27.9% claim is less affected: Table 2 compares to Diffusion Forcing (train-only), so that part of the abstract survives this concern.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Stream Forcing, a curriculum training framework for streaming video diffusion models. It reformulates per-frame noise-level sampling during training as a frame-indexed stochastic process with Logit-Normal marginals, parameterized by location, scale, and temporal correlation. The method constructs a training trajectory from independent sampling (as in Diffusion Forcing) to inference-aligned progressive sampling (as in rolling/AR diffusion), enforced by a joint calibration algorithm for mode and scale parameters and a Gaussian-copula-based temporal correlative sampling scheme. Experiments on UCF-101, Taichi-HD, and nuScenes report FVD improvements, including zero-shot long-horizon extrapolation to 128 frames.","tokens_in":16087,"tokens_out":9726,"duration_ms":80886,"significance":"The core idea of interpolating between independent and progressive noise-level schedules is timely and potentially useful for streaming video world models. The derivation of the mode equation in Appendix A.4 is correct, and the ablation studies (Tables 4-7) show that each proposed component contributes to the reported FVD. However, the headline 36.6% UCF-101 improvement is computed against a baseline (FAR*) trained on the train+test split, while the proposed method is trained on the training split only. The 27.9% long-horizon gain is more credible because it is based on train-only baselines. Thus, the conceptual contribution is defensible but the central quantitative claim needs correction.","major_comments":[{"comment":"The abstract and Section 4.1.1 report a 36.6% FVD improvement on UCF-101 (from 279.0 to 177.0), but this is computed against FAR*, which Table 1 marks with * meaning 'trained on both the training and test splits.' Stream Forcing is trained on the training split only. The text claims this comparison is 'under the same evaluation setting,' which the table contradicts. Against the best train-only baseline, AR-Diffusion† (FVD 181.9), the improvement of Ours (177.0) is (181.9 - 177.0)/181.9 ≈ 2.7%, and the improvement of Ours† (146.9) is ≈ 19.2%. Because the abstract, introduction, and experiments all lead with the 36.6% number, the central empirical claim is not supported by the reported comparisons. Please recompute headline comparisons against train-only baselines, or explicitly justify and qualify the comparison to train+test methods.","section":"Section 4.1.1 / Table 1"}],"minor_comments":[{"comment":"The curriculum is discretized into 10 uniformly-spaced configurations; the paper does not justify this number or provide a sensitivity analysis, so please add a discussion or ablation of the number of curriculum points.","section":"Section 3.3"},{"comment":"The joint calibration grid search does not specify the number of grid points N, the upper bound of the search range for sigma, or the stopping tolerance; please provide these details for reproducibility.","section":"Algorithm 1"},{"comment":"The ablation table rows are ambiguous in formatting; please show all three constraint indicators (TS, DCC, IFC) for each row so the reader can see which constraint is removed.","section":"Table 4"},{"comment":"The caption does not annotate which baselines are trained on the training split only versus train+test, unlike Table 1; please add matching footnotes to allow verification of the reported zero-shot improvements.","section":"Table 2"},{"comment":"The AR(1) process does not specify the distribution of the initial value z_0; please state that z_0 is standard normal so that all marginals are standard normal as claimed.","section":"Equation 10"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is sound and the internal ablations are consistent, but the headline quantitative claim is overstated because of the train+test baseline in Table 1. The authors should be asked to recompute the headline comparison against train-only baselines and to clarify the evaluation protocol. The long-horizon and nuScenes results are more convincing and suggest the method has real value."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: the training-recipe idea is genuinely worth your attention, but the headline number is not what it appears to be. The 36.6% FVD improvement over 'state-of-the-art' on UCF-101 is computed against FAR*, which the paper's own Table 1 marks as trained on both train and test splits. Against the best train-only baseline (AR-Diffusion†, 181.9), the gain is about 2.7%. The abstract and intro lead with the 36.6% number, and that is a real problem.\n\nWhat is actually new: the paper rethinks training-time noise-level sampling as a frame-indexed stochastic process with Logit-normal marginals and a copula-induced temporal correlation, then builds a curriculum from independent sampling (Diffusion Forcing-style) to progressive, inference-aligned sampling. That unification is not in the cited prior work, and it's a sensible way to frame the trade-off between coverage and consistency. The mode equation proof in Appendix A.4 is straightforward and correct. The ablations (Tables 4-7) consistently show that each constraint—smooth mode transitions, per-frame density matching, and inter-frame correlation—pulls FVD down, which is evidence the design is doing something real. The zero-shot long-horizon result (27.9% on UCF-101, Table 2) compares to Diffusion Forcing, which is train-only, so that claim survives the baseline criticism.\n\nSoft spots, in order of severity. First, the evaluation-comparability problem above. It's not a math error; it's a claim about what baseline you're beating. The paper should re-report the headline against train-only state-of-the-art and label the FAR* comparison as a separate, clearly-marked row. Second, there are no error bars or multiple seeds; FVD gaps around 2-5% can be noise at this scale. Third, key hyperparameters (ρ_max, the 2:1 phase ratio) are selected via ablation on the eval benchmark, which doesn't invalidate the method but does soften the 'predictive' strength of the numbers. Fourth, the main table reports 177 FVD after 600k steps, but the ablation full model gets 334.4 at 450k steps; that 150k-step gap is large and not discussed. Maybe it's just training budget, but it deserves a sentence.\n\nThe paper is for people working on streaming video diffusion, world models, or long-horizon generation. The framework is plausible, the math is clean, and the flaws are fixable. I'd send it to peer review, but with a clear instruction to the authors: re-run the claims against train-only baselines, add variance, and release code. As is, the abstract overclaims, and reviewers should not let that slide.","headline":"Solid training-curriculum idea with a clean proof and consistent ablations, but the headline 36.6% FVD claim rests on a train+test baseline; the honest same-setting gain is about 2.7%.","tokens_in":16666,"tokens_out":3924,"would_cite":true,"duration_ms":32711,"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":"Stream Forcing unifies the two dominant training strategies for streaming video diffusion by interpolating their noise-level schedules along a smooth curriculum, and reports that this unified training improves video quality and zero-shot…","keywords":["streaming video generation","video diffusion models","curriculum training","noise-level sampling","Logit-Normal distribution","Gaussian copula","train-inference consistency","Fréchet Video Distance"],"falsifier":"Retrain FAR* on the UCF-101 training split only and evaluate under the exact 16-frame unconditional protocol: if its FVD falls to or below 177.0, the claimed 36.6% improvement disappears. A simpler check is to compare directly with the train-only AR-Diffusion† result of 181.9, where the margin is about 2.7% rather than 36.6%.","tokens_in":1660,"feed_emoji":"🎬","tokens_out":3424,"duration_ms":114405,"temperature":0.7,"pith_summary":"Streaming video diffusion models face a conflict: training usually draws noise levels independently per frame, which gives broad coverage but ignores the ordered denoising used at inference, while training that mimics the inference schedule improves consistency but narrows the noise levels seen. Stream Forcing reformulates training-time noise-level sampling as a frame-indexed stochastic process with Logit-Normal marginals and an inter-frame correlation parameter, making independent and progressive sampling the two endpoints of one family. Training walks a smooth curriculum from the independent endpoint to the inference-aligned endpoint using joint calibration and Gaussian-Copula-based correlated sampling. The authors report that this unified trajectory improves the Fréchet Video Distance (FVD) by 36.6% on UCF-101 and 4.7% on Taichi-HD, and enables zero-shot extrapolation to 128-frame videos with 27.9% and 10.9% FVD improvements. If these results hold, the method offers a general way to align diffusion training with causal streaming inference without sacrificing training coverage.","feed_headline":"Stream Forcing eases the training-inference gap in streaming video","feed_subtitle":"A curriculum from independent to progressive noise sampling cuts FVD and extends zero-shot to 128-frame clips.","key_machinery":"The central object is the frame-indexed stochastic process over noise levels, with Logit-Normal marginals—a distribution on $(0,1)$ obtained by applying the logistic sigmoid to a Gaussian random variable—parameterized by per-frame $(\\mu_t,\\sigma_t)$ and an inter-frame correlation $\\rho$. The mode equation $\\psi(\\zeta,\\sigma)=\\logit(\\zeta)+\\sigma^2(1-2\\zeta)$ converts a desired marginal mode $\\zeta$ into a location parameter, and a joint calibration algorithm grid-searches $\\sigma$ so that all frames share a common peak density. A Gaussian Copula built from an AR(1) process, $z_t=\\rho z_{t-1}+\\sqrt{1-\\rho^2}\\,\\epsilon_t$, generates correlated noise levels while preserving each frame's prescribed marginal. These pieces implement the curriculum that evolves the sampling configuration from independent to inference-consistent.","core_discovery":"The paper argues that the train-inference mismatch in streaming video diffusion is not a binary choice. It reformulates training-time noise-level sampling as a frame-indexed stochastic process whose marginals are Logit-Normal, controlled per frame by a location $\\mu_t$ and scale $\\sigma_t$ and across frames by a correlation $\\rho$. In this space, independent sampling ($\\mu_t=\\mu_0$, $\\sigma_t=\\sigma_0$, $\\rho=0$) and progressive sampling ($\\mu_t\\to \\logit(t/T)$, $\\sigma_t\\to 0$, $\\rho\\to 1$) are endpoints of one continuous family. Stream Forcing trains along a curriculum that interpolates between these endpoints, starting with independent noise for broad coverage, passing through calibrated intermediate configurations, and finishing with inference-consistent sampling. Joint calibration keeps per-frame mode densities balanced, and a Gaussian Copula realizes the prescribed inter-frame correlation. The central claim is that this unified trajectory produces higher-quality streaming video than either endpoint alone and transfers zero-shot to longer horizons.","pith_inferences":["The same stochastic-process interpolation could apply to other causal sequence diffusion tasks—audio, control, or sensor streams—where an autoregressive generation order conflicts with full-sequence training.","The grid-search calibration could likely be replaced by a closed-form dispersion-matching step, which would make the curriculum easier to port to new datasets; the paper's own entropy ablations suggest mode-density matching is the active ingredient.","The headline margin depends on evaluation protocol: against the train-only AR-Diffusion† baseline the UCF-101 gap is about 2.7%, so a strictly train-only rerun of FAR* would clarify how much of the gain is due to the curriculum."],"forward_implications":["Streaming video diffusion can be trained with broad noise coverage while still matching the ordered denoising used at inference, so practitioners no longer have to choose between the two.","The same curriculum transfers to a new domain: on nuScenes driving videos, Stream Forcing improves both FID and FVD over existing driving-world-model baselines.","Models trained on 16-frame clips can extrapolate to 128-frame streams without extra long-video training, with reported FVD gains of 27.9% on UCF-101 and 10.9% on Taichi-HD.","The three-stage schedule introduces a tunable balance between coverage and consistency; the ablation finds a 2:1 ratio of independent to curriculum training performs best."],"supporting_citations":[{"why":"Supplies the Logit-Normal density used as the marginal distribution for per-frame noise-level sampling.","marker":"[1]"},{"why":"Defines the independent-sampling endpoint and serves as a main comparative baseline.","marker":"[6]"},{"why":"Defines the progressive-sampling endpoint whose inference-aligned schedule the training trajectory targets.","marker":"[35]"},{"why":"Provides the progressive-sampling baseline and the controlled-comparison setting used for the train-only FVD numbers.","marker":"[41]"},{"why":"Supplies the Gaussian Copula construction that adds inter-frame correlation without altering the marginals.","marker":"[33]"},{"why":"Is the state-of-the-art baseline whose FVD score anchors the headline 36.6% improvement.","marker":"[13]"},{"why":"Provides the DFoT DiT backbone used in the UCF-101 and Taichi-HD experiments.","marker":"[39]"},{"why":"Is the UCF-101 benchmark dataset on which the main quality and long-horizon results are reported.","marker":"[40]"}],"fun_headline_variants":["Stream Forcing: one training path to fix streaming video's mismatch","Unified noise curriculum improves streaming video by 36.6% FVD","From independent to consistent: Stream Forcing's robust trajectory","Stream Forcing enables zero-shot long video with calibrated sampling"],"cache_read_input_tokens":18816,"weakest_assumption_plain":"The headline quality gain assumes it is fair to compare a model trained only on the training split with baselines trained on both training and test splits; the table itself marks the main baseline that way, so the reported margin depends on that comparability.","fun_headline_variants_meta":{"raw":{"variants":["Stream Forcing: one training path to fix streaming video's mismatch","Unified noise curriculum improves streaming video by 36.6% FVD","From independent to consistent: Stream Forcing's robust trajectory","Stream Forcing enables zero-shot long video with calibrated sampling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000281,"raw_usage":{"total_tokens":1671,"prompt_tokens":962,"completion_tokens":709,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":636}},"tokens_in":578,"tokens_out":709,"duration_ms":7009,"temperature":1.0,"reasoning_tokens":636,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:20:32.948038+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain FAR* on the UCF-101 training split only and evaluate under the exact 16-frame unconditional protocol: if its FVD falls to or below 177.0, the claimed 36.6% improvement disappears. A simpler check is to compare directly with the train-only AR-Diffusion† result of 181.9, where the margin is about 2.7% rather than 36.6%.","supporting_citations":[{"cited_title":"Logistic-normal distribu- tions: Some properties and uses.Biometrika, 67(2):261–272,","cited_arxiv_id":null,"evidence_quote":"Supplies the Logit-Normal density used as the marginal distribution for per-frame noise-level sampling."},{"cited_title":"Diffusion forcing: Next-token prediction meets full-sequence diffu- sion.Advances in Neural Information Processing Systems, 37:24081–24125, 2024","cited_arxiv_id":null,"evidence_quote":"Defines the independent-sampling endpoint and serves as a main comparative baseline."},{"cited_title":"Rolling diffusion models","cited_arxiv_id":null,"evidence_quote":"Defines the progressive-sampling endpoint whose inference-aligned schedule the training trajectory targets."},{"cited_title":"Ar-diffusion: Asynchronous video genera- tion with auto-regressive diffusion","cited_arxiv_id":null,"evidence_quote":"Provides the progressive-sampling baseline and the controlled-comparison setting used for the train-only FVD numbers."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian Copula construction that adds inter-frame correlation without altering the marginals."}],"review_version":1}