{"id":"e7b02f38-dcd7-4f9a-b26e-b5cf85d17d94","arxiv_id":"2501.05484","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"GLC-Diffusion extends short-clip video diffusion models to long videos via global-local collaborative denoising, noise reinitialization, and motion-consistency refinement, improving coherence and fidelity at 3x and 6x longer lengths.","lead":"GLC-Diffusion is a training-free method that makes existing AI video generators produce longer, more coherent videos, tested up to six times their original length. It combines global and local denoising paths, noise reinitialization, and motion-consistency refinement, and could make long AI video generation practical without costly retraining.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Superiority over FreeNoise rests on <1.3-point VBench margins from a single run, with four hyperparameters selected on the same 200 prompts; the claim needs held-out, error-barred evaluation.","rationale":"The reader's CONDITIONAL verdict is appropriate, but the specific weakest assumption (Eq. 8-9 derivation) is not the most load-bearing. The empirical superiority claim rests on a single table with small margins, no error bars, and hyperparameters selected on the same evaluation set. A held-out or multi-seed rerun would directly settle whether the 1-2 point VBench advantages are real or selection noise. I therefore keep the reader's CONDITIONAL verdict unchanged rather than moving to ACCEPT or REJECT. I agree partially because the reader also flags the small margins and missing error bars; my concern differs in identifying the same-set hyperparameter selection as the decisive point, not the optimization derivation.","tokens_in":29760,"tokens_out":3972,"duration_ms":40062,"concrete_test":"Fix all hyperparameters to the appendix defaults, select 200 VBench prompts disjoint from the tuning set (or, failing that, 3 fixed seeds on the original 200), and re-run Table 1 for Direct Sampling, FreeLong, GenL, FreeNoise, and GLC-Diffusion. Report per-video means with bootstrap 95% CIs. If GLC-Diffusion no longer beats FreeNoise/GenL on the majority of the five metrics at both 3x and 6x, or if the Imaging margins fall inside the CIs, the superiority claim should be downgraded to 'competitive' rather than 'superior.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's claim that GLC-Diffusion produces long videos 'superior to previous approaches' is carried almost entirely by Table 1. The largest advantage over the best baseline FreeNoise is +1.22 Imaging at 3x and +1.12 at 6x; Motion and Flicker gains are +0.12/+0.26 and +0.23/+0.46. No per-video distributions, seeds, or confidence intervals are reported, so those differences may be within run-to-run noise. More importantly, the appendix shows that all four central hyperparameters — gamma0 in GLCD (Table 3), lambda in ABAM (Table 4), lambda_f in VMCR (Table 5), and omega_motion (Table 6) — were selected by sweeping the same VBench metrics on the same 200 prompts used for Table 1. No held-out split is described. This is a selection-on-the-test-set protocol: the reported advantage can reflect fitting noise rather than a robust method advantage. The Eq. 8-to-Eq. 9 derivation gap noted by the reader is real but secondary: Eq. 9 is an asserted convex blend, not the least-squares solution of Eq. 8, yet the method could still be empirically effective. The missing held-out, error-barred comparison is the load-bearing issue for the paper's central superiority claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GLC-Diffusion, a tuning-free, plug-and-play method for extending short video diffusion models to generate longer videos. The method combines two denoising paths: a global path using dilated sampling to capture long-range dependencies and a local path using randomly shifted overlapping clips for local temporal coherence, fused through an annealing coefficient. It also introduces an anchor-based attention modulation (ABAM), a noise reinitialization strategy that shuffles local noise and fuses low/high frequency components, and a video motion consistency refinement (VMCR) module that aligns motion vectors through pixel-wise and frequency-wise losses. Experiments on CogVideoX generate videos at 3x and 6x the original length and report VBench scores over 200 prompts, claiming superiority over Direct Sampling, Gen-L-Video, FreeNoise, and FreeLong. Ablations and hyperparameter studies are given in the appendix.","tokens_in":30042,"tokens_out":4209,"duration_ms":43042,"significance":"If the results are robust, the contribution is practically useful: a training-free extension that scales a 49-frame model to hundreds or thousands of frames while preserving content consistency and temporal coherence is of clear interest to the video generation community. The modular design and the inclusion of ablations for each component are strengths, and the method's compatibility with an existing model such as CogVideoX lends credibility to the plug-and-play claim. However, the central claim of superiority over prior work is currently supported only by small margins on a single evaluation run, with hyperparameters selected on the same 200 prompts used for evaluation. This is a selection-on-the-test-set protocol that, combined with the absence of error bars, makes the reported advantage potentially attributable to fitting evaluation noise rather than a robust method advantage.","major_comments":[{"comment":"The headline claim that GLC-Diffusion is superior to prior methods rests on very small margins over FreeNoise. In Table 1, the largest advantage at 3x is +1.22 in Imaging, while Motion, Flicker, and Background gains are only +0.12, +0.23, and +0.66, respectively; at 6x the Motion and Flicker gains are +0.26 and +0.46. No seeds, per-video distributions, confidence intervals, or significance tests are reported. More importantly, the appendix shows that all four central hyperparameters—gamma0 in GLCD (Table 3), lambda in ABAM (Table 4), lambda_f in VMCR (Table 5), and omega_motion (Table 6)—were selected by sweeping VBench metrics on the same 200 prompts used to produce Table 1. No held-out split is described. Under this protocol, the reported margins are consistent with selection on evaluation noise. The authors should provide multiple seeded runs with confidence intervals, use a held-out split for any hyperparameter selection, or perform significance tests before claiming superiority.","section":"Section 4.2, Table 1; Appendix Tables 3-6"},{"comment":"Algorithm 1 in the supplementary material is inconsistent with Eq. (9) and with the described method. Eq. (9) defines z_{t-1} = gamma * T_global(Z_global) + (1-gamma) * T_local(Z_local), and the text states that the global and local paths are fused with the annealing coefficient. However, Algorithm 1 line 17 writes only z_{t-1} = gamma * T_global(Z_global), with a comment saying \"Combine global and local paths,\" and the local term is absent. If the implementation matches Algorithm 1, the local path is not used at all, which contradicts the ablation study showing that removing the local path substantially degrades results. This must be corrected or clarified, as it is load-bearing for reproducibility.","section":"Algorithm 1, line 17"},{"comment":"The derivation from the weighted least-squares objective in Eq. (8) to the closed-form blend in Eq. (9) is asserted rather than proven. In general, the solution of Eq. (8) depends on the overlap structure and the weight matrices W_k; it is not generically equal to a simple convex combination of independent global and local reconstructions, gamma * T_global + (1-gamma) * T_local. Invoking the \"manifold hypothesis\" does not by itself establish this equality. If the optimization framing cannot be substantiated, the method should be presented as a heuristic weighted fusion of two denoising paths. The empirical results may still stand, but the current statement overclaims the theoretical grounding of Eq. (9).","section":"Section 3.2, Eqs. (8)-(9)"},{"comment":"The ablation study also lacks error bars or repeated runs, which matters for the interpretation of the results. In particular, removing Noise Reinitialization changes Imaging from 69.86 to 69.71, a 0.15-point difference that is almost certainly within run-to-run noise, yet the text describes the noise reinitialization module as having a significant impact. The same issue applies to the hyperparameter sweeps in Tables 3-6, where the selected values are often separated by differences comparable to the margins over baselines in Table 1. The authors should temper the language or provide statistical support for the claimed improvements.","section":"Section 4.4, Table 2"}],"minor_comments":[{"comment":"The paper inconsistently states that CogVideoX produces 48 frames (Introduction) and 49 frames (Section 4.1 and Figure 1). Please reconcile these numbers.","section":"Introduction, Section 4.1, Figure 1"},{"comment":"In Eq. (4), the variable j is described as \"the j-th video clip\" but it is used as a frame offset within a clip (j = 0, ..., L-1). This is confusing; j should be described as the frame index inside the clip, and the mapping between clip index i and sampled frames should be stated more precisely.","section":"Section 3.2, Eq. (4)"},{"comment":"The hyperparameter section states that gamma0 = 0.005 and beta = 0.0005, but it does not give the total number of timesteps T used for DDIM sampling. Since gamma = gamma0 * exp(beta * t) depends on t, the effective range of gamma cannot be assessed without T. Please state the schedule and the resulting gamma at the final timestep.","section":"Supplementary, Section 7"}],"recommendation":"major_revision","confidential_remarks":"The paper is a plausible empirical contribution, but the evaluation protocol needs substantial strengthening before the superiority claim can be accepted. The authors should be asked to provide seeded runs with error bars, a held-out split for hyperparameter selection, and a resolution of the Algorithm 1 inconsistency. If these cannot be provided, the abstract and conclusion should be scaled back to claim competitiveness rather than clear superiority."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this is a solid, plug-and-play extension for tuning-free long video generation. The assembly of global dilated sampling, local random shifting, anchor-based attention, noise reinitialization with frequency fusion, and the VMCR gradient refinement is new relative to FreeNoise and FreeLong. Each ingredient has a clear precursor, but the combination makes sense and the ablations show every component earns its keep. The qualitative results look plausible and the writing is honest, including a limitation note about complex motion. This is a real contribution, not a paradigm shift.\n\nThe soft spots are mostly about evidence, not method. The quantitative story rests on Table 1, and the margins over the best baseline are small: +1.22 Imaging at 3x, +1.12 at 6x, and much less on Motion and Flicker. No error bars, seeds, or per-video distributions are reported, so those differences might be within run-to-run noise. More importantly, the appendix shows that all four central hyperparameters (gamma0, lambda, lambda_f, omega_motion) were selected by sweeping the same VBench metrics on the same 200 prompts used for the main comparison. No held-out split is described. That is selection-on-the-test-set, and it means the reported advantage could reflect fitting noise rather than a robust method advantage. This is the load-bearing issue for the paper's central claim.\n\nThe Eq. 8-to-Eq. 9 step is also asserted rather than derived: Eq. 9 is a convex blend, not obviously the least-squares solution of Eq. 8. The reader's stress-test note is right that this is secondary, since the method could still work empirically, but it should be fixed or the framing softened.\n\nFor a researcher working on tuning-free long video generation, this paper is worth reading and engaging with. It deserves a serious referee, but the current version should not be accepted as-is. I would ask for held-out, error-barred evaluation, ideally with the hyperparameters frozen on a validation split, and for code or videos to back up the claims. The method itself is plausible enough that the extra evidence is worth the referee time.\n\nMy bottom line: conditional acceptance with revisions, and I'd bring it to a reading group as a good example of both a useful method and a cautionary tale about evaluation protocol.","headline":"A sensible, well-ablated plug-and-play method for long video generation; the superiority claim is not yet proven because the margins over FreeNoise are small and the hyperparameters were tuned on the same VBench prompts used for evaluation.","tokens_in":30586,"tokens_out":1662,"would_cite":true,"duration_ms":18274,"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":"This paper claims that a tuning-free plugin can make a 49-frame video diffusion model generate coherent videos of over 1,000 frames by fusing global and local denoising paths.","keywords":["long video generation","tuning-free","video diffusion models","global-local collaborative denoising","noise reinitialization","motion consistency refinement","temporal coherence","VBench"],"falsifier":"Compute the argmin of Eq. 8 directly on a real denoising step, keeping the global and local clips fixed, and compare it with the Eq. 9 weighted average; if the two differ, the paper's stated optimization is not actually being solved by the implemented fusion. A complementary external check is to replace the closed-form blend with the numerically solved least-squares reconstruction and rerun the same evaluation protocol; equal or worse scores would indicate that the blend, not the optimization, is the working ingredient.","tokens_in":29556,"feed_emoji":"🎬","tokens_out":6301,"duration_ms":56963,"temperature":0.7,"pith_summary":"This paper claims that long-video generation can be improved without any retraining by splitting each denoising step into a global path and a local path and fusing the two. The resulting method, GLC-Diffusion, is a plug-in for an existing short-clip video diffusion model, extending it from 49 frames to hundreds or more than a thousand frames while keeping content and motion consistent. If the claim holds, pretrained video diffusion models can be upgraded to much longer outputs by adding three modules, collaborative denoising, noise reinitialization, and motion refinement, rather than retraining on long-video data. The paper reports the best scores among four tuning-free baselines on all five consistency and fidelity metrics it evaluates.","feed_headline":"A plug-in stretches 49-frame video models past 1,000 frames","feed_subtitle":"Global and local denoising paths plus motion refinement beat four baselines on long video tests.","key_machinery":"The load-bearing object is Global-Local Collaborative Denoising (GLCD), framed as a single weighted least-squares problem over all global and local clips (Eq. 8). Global Dilated Sampling builds clips by taking frames $d$ apart, so the model sees long-range structure; Local Random Shifting Sampling builds overlapping clips with start indices shifted randomly at each timestep, so seams and flicker get corrected from multiple temporal viewpoints. The two paths are merged by the annealing blend $z_{t-1} = \\gamma\\,T_{\\mathrm{global}}(Z^{\\mathrm{global}}_{t-1}) + (1-\\gamma)\\,T_{\\mathrm{local}}(Z^{\\mathrm{local}}_{t-1})$, where $T$ is the clip-merge operation of Eq. 3 and $\\gamma$ grows with timestep. Around this core sit Noise Reinitialization (local noise shuffle plus frequency fusion of the initial latents), the Anchor-Based Attention Mechanism (ABAM) that injects the first clip's keys and values into other clips' attention, and Video Motion Consistency Refinement (VMCR), a gradient step on a pixel-plus-frequency motion loss.","core_discovery":"In the paper's own telling, GLC-Diffusion is a tuning-free recipe that turns a short-clip video diffusion model into a long-video generator. The denoising trajectory is split into a global path and a local path; global dilated sampling holds the overall scene and narrative steady, while local clips with random temporal shifts smooth frame-to-frame motion. The two paths are fused at every timestep by a weighted blend with an annealing coefficient, the initial noise is reinitialized by shuffling local noise and recombining low frequencies with fresh high-frequency noise, and a gradient-based motion refinement aligns adjacent frames in both pixel and Fourier domains. On 200 prompts from the VBench benchmark, the paper reports the highest scores among Direct Sampling, Gen-L-Video, FreeNoise, and FreeLong on all five reported consistency and fidelity metrics at both $3\\times$ and $6\\times$ lengths, and demonstrates a video of more than 1,000 frames from a model trained on 49 frames.","pith_inferences":["Going beyond the paper, the two-path fusion only needs clip projections and denoising calls, so it could likely be dropped into other temporal-attention video models.","A direct numerical solution of Eq. 8 would test whether the claimed optimum matters or whether the weighted average alone is the working ingredient.","The VMCR frequency-domain motion alignment could be used as a post-hoc smoothing regularizer on any video sampling pipeline.","The noise-reinitialization idea offers a low-cost way to add output diversity to existing short-video models."],"forward_implications":["A pretrained 49-frame text-to-video model can generate 294-frame and longer videos without additional training, with higher reported scores than four tuning-free baselines.","Long-range content consistency and local frame-to-frame coherence are both improved by combining dilated global sampling with randomly shifted overlapping local clips.","Noise reinitialization through local shuffling plus frequency fusion increases visual diversity while keeping overall content stable.","The VMCR gradient step improves motion smoothness and flicker metrics across the whole generated video.","The same recipe scales a 49-frame model to more than 1,000 frames, roughly 25 times the trained clip length."],"supporting_citations":[{"why":"Supplies the weighted least-squares fusion of denoising trajectories and the manifold hypothesis used to justify the closed-form blend in Eq. 9.","marker":"[1]"},{"why":"Provides the deterministic DDIM sampling update used as the base denoising step for all paths.","marker":"[34]"},{"why":"Introduces the clip projection and merging approach (Eqs. 2-3) that GLCD extends and compares against as a baseline.","marker":"[35]"},{"why":"Is the FreeNoise baseline and the source of local noise shuffling used in the Noise Reinitialization strategy.","marker":"[31]"},{"why":"Is the FreeLong baseline that fuses global low-frequency and local high-frequency features, the main comparative context for global-local balancing.","marker":"[22]"},{"why":"Is the pretrained 49-frame latent video diffusion model used as the backbone for all experiments.","marker":"[43]"},{"why":"Provides the VBench benchmark and the five evaluation metrics used for quantitative comparisons.","marker":"[13]"},{"why":"Inspires the initialization-frequency fusion component of the Noise Reinitialization strategy.","marker":"[40]"}],"fun_headline_variants":["Short-clip models stretch to 1,000 frames with tuning-free method","Global-local denoising paths keep long videos coherent without tuning","49-frame diffusion seeds 1,000-frame videos via collaborative paths","Noise shuffle and motion refinement make short models go long","GLC-Diffusion: tune-free long video from any short-clip model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's theoretical story depends on the claim that the weighted least-squares problem in Eq. 8 has the convex blend in Eq. 9 as its exact solution, asserted via the manifold hypothesis without a derivation or numerical check.","fun_headline_variants_meta":{"raw":{"variants":["Short-clip models stretch to 1,000 frames with tuning-free method","Global-local denoising paths keep long videos coherent without tuning","49-frame diffusion seeds 1,000-frame videos via collaborative paths","Noise shuffle and motion refinement make short models go long","GLC-Diffusion: tune-free long video from any short-clip model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000158,"raw_usage":{"total_tokens":1212,"prompt_tokens":920,"completion_tokens":292,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":199}},"tokens_in":536,"tokens_out":292,"duration_ms":3317,"temperature":1.0,"reasoning_tokens":199,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:36:14.912342+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the argmin of Eq. 8 directly on a real denoising step, keeping the global and local clips fixed, and compare it with the Eq. 9 weighted average; if the two differ, the paper's stated optimization is not actually being solved by the implemented fusion. A complementary external check is to replace the closed-form blend with the numerically solved least-squares reconstruction and rerun the same evaluation protocol; equal or worse scores would indicate that the blend, not the optimization, is the working ingredient.","supporting_citations":[{"cited_title":"MultiDiffusion: Fusing diffusion paths for controlled image generation","cited_arxiv_id":null,"evidence_quote":"Supplies the weighted least-squares fusion of denoising trajectories and the manifold hypothesis used to justify the closed-form blend in Eq. 9."},{"cited_title":"Freenoise: Tuning- free longer video diffusion via noise rescheduling","cited_arxiv_id":null,"evidence_quote":"Is the FreeNoise baseline and the source of local noise shuffling used in the Noise Reinitialization strategy."},{"cited_title":"Freeinit: Bridging initialization gap in video dif- fusion models","cited_arxiv_id":null,"evidence_quote":"Inspires the initialization-frequency fusion component of the Noise Reinitialization strategy."}],"review_version":1}