{"id":"6d6c7f08-7917-4e4e-a2fd-77b535f9c2ff","arxiv_id":"2501.02741","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Long videos are generated by denoising short latent segments with a shifting offset between steps, and the approach reports higher VBench average scores than four training-free baselines.","lead":"A training-free method called Brick-Diffusion makes long videos by denoising short latent segments with a shifting offset between steps, like staggered bricks in a wall. The paper reports higher VBench scores than four existing training-free long-video methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (6)'s independent-segment denoising assumes each slice has training-like statistics, but shifted regrouping may only relocate boundary discontinuities; this is untested.","rationale":"I agree with the reader that the weakest link is the assumption in Eq. (6) that independent segment-wise denoising approximates joint denoising. The paper provides no theoretical or empirical validation of this approximation beyond an ablation showing that stride>0 helps. The ablation is necessary but not sufficient: stride>0 would also help if the method merely spreads artifacts more evenly, making them harder to see in aggregate metrics. The concern is load-bearing because the entire method is a heuristic modification of the diffusion sampling loop; if the slice statistics deviate from training distribution, the noise predictions become biased, and errors can compound over many steps. The concrete test I propose directly probes the hypothesized boundary artifact by comparing frames that are frequently at slice boundaries against those that are not. This is a targeted, low-cost check that would either confirm the method's robustness or expose a systematic flaw. The reader's CONDITIONAL verdict remains appropriate: the paper has a plausible idea and a credible ablation, but the central claim of superiority needs this additional evidence. I do not see grounds to reject outright, since the method may work well in practice despite the lack of formal justification; nor should it be accepted without the boundary-artifact check and ideally multi-seed error bars. Therefore I recommend no change to the reader's verdict.","tokens_in":7233,"tokens_out":5926,"duration_ms":66754,"concrete_test":"Using the generated 128-frame videos (or generating new ones), compute for each frame the number of times during the last 10 DDIM steps it fell within 1 frame of a slice boundary under the stride schedule of Eq. (3). Split frames into a high-exposure group (>5 boundary events) and a low-exposure group (<2 events). For each group, compute mean warped-frame error using RAFT (or CLIP temporal consistency) over the 825 videos. If high-exposure frames show significantly larger error (e.g., >15% relative increase), the shifting stride fails to repair boundary discontinuities and the central fidelity claim is weakened. If the two groups are statistically indistinguishable, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The method's core validity rests on Eq. (6) in Section II-B, where each f-frame segment is denoised as if it were an independent sample from the pre-trained model's training distribution at noise level t. This is justified only if every slice of the long latent has the same marginal and cross-frame statistics as q(z_t | x) for some clean video x. However, after the first denoising step, the latent is a concatenation of independently denoised blocks. When the stride shifts the slicing grid, a new segment is formed by frames from different previous blocks that have no shared denoising history. The model's temporal attention then observes an artificial discontinuity in noise correlation exactly at the boundaries of the old blocks. The shifting stride moves these boundaries by one frame each step, but it does not merge the histories; it merely redistributes the discontinuity across frames and timesteps. Thus the approximation error is not repaired, only smeared. The paper's only evidence for 'communication' between segments is the stride ablation (Table II), which compares stride=0 versus non-zero; it does not measure residual boundary artifacts or demonstrate convergence toward joint denoising. VBench's aggregate metrics (subject consistency, dynamic degree, aesthetic quality, text consistency) are not sensitive to localized boundary flicker or temporal discontinuities, so the central claim of high fidelity may rest on an unvalidated heuristic.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Brick-Diffusion proposes a training-free inference procedure that generates long videos by slicing a latent into f-frame segments, denoising each segment independently with a pretrained short-video diffusion model, and shifting the slicing grid by a stride at every denoising step. The authors argue that this 'brick-to-wall' shifting lets frames from different segments communicate, and they evaluate the method with VideoCrafter2 on the VBench benchmark for 128-frame videos. In their quantitative results, the method attains an average score of 81.70, ahead of FIFO-Diffusion's 79.12, driven mainly by a large dynamic-degree advantage and a small text-consistency advantage, while FreeNoise remains best on subject consistency. A stride ablation shows that non-zero strides greatly improve over stride 0, with s=1 selected as the best.","tokens_in":7471,"tokens_out":8982,"duration_ms":87069,"significance":"Brick-Diffusion is a simple and potentially useful contribution: if its reported gains are reproducible, it offers a parallelizable, fine-tuning-free way to extend existing short-video diffusion models to longer videos, and the paper makes this claim falsifiable through the stride ablation. The algorithm is described clearly enough to reimplement, and the evaluation against four training-free baselines on a standard benchmark is a strength. However, the significance is currently limited by the lack of statistical robustness in the headline comparison, the selection of the stride on the same test set used for reporting, and the absence of direct evidence that the independent-segment approximation in Eq. (6) preserves the joint statistics required for temporal consistency. These issues are addressable and do not invalidate the core idea, but they preclude acceptance of the superiority claims as stated.","major_comments":[{"comment":"The headline comparison reports a single numeric value per metric and method with no error bars, confidence intervals, or number of independent runs. The paper states that 825 videos are generated per method, but it does not say whether these come from one seed per prompt or from repeated draws; if they are single draws, the reported 2.58-point average margin over FIFO-Diffusion and the 9.16-point dynamic-degree margin could change substantially under a different initial noise. Since FIFO-Diffusion actually leads on subject consistency (93.91 vs 93.00), the 'superiority' claim needs a statistical basis. Please report means and standard errors over at least three seeds, or otherwise justify why a single run is representative.","section":"Table I and Section III-D"},{"comment":"The stride s=1 is selected as the configuration with the highest average score on the same VBench prompts used for the final comparison against baselines. This is an instance of tuning on the test set, and it makes the reported advantage over the baselines optimistic because the baselines receive no comparably tuned hyperparameter. The authors should pre-register the stride, choose it on a separate validation set, or report all stride values alongside the baseline comparison and discuss the selection bias. The near-flat results for s=3,5,7,9 suggest robustness, but the selection protocol must still be stated.","section":"Table II and Section III-C"},{"comment":"The core of the method is the assumption that denoising f-frame segments independently at each timestep, with the grid shifted between steps, is a valid approximation of the joint denoising the pretrained model would perform on the whole clip. The paper does not provide direct evidence for this assumption: the stride ablation in Table II only shows that non-zero stride beats concatenation, while VBench's aggregate metrics (DINOv2-based subject consistency, RAFT-based dynamic degree, aesthetic score, InternVid text-video similarity) are unlikely to be sensitive to localized flicker or discontinuity exactly at the old segment boundaries. Please add a diagnostic that isolates boundary artifacts, for example comparing per-frame metrics near versus far from previous slice boundaries, testing a control in which segment order is randomly permuted between steps, or comparing the final latent against a jointly denoised short-clip reference. Without such evidence, the 'communication between segments' mechanism remains a heuristic.","section":"Section II-B, Eq. (6)"},{"comment":"The qualitative evaluation consists of a single text prompt ('a cute raccoon playing guitar in a boat on the ocean'). A single example cannot support the claim of consistently high-fidelity output, especially when the advertised differences are temporal (flicker, object deformation, disappearance). Please include a supplementary video set or a small perceptual study (e.g., pairwise preference among methods over several prompts), and show frames from at least a few diverse prompts.","section":"Section III-D, Fig. 3"}],"minor_comments":[{"comment":"Reference [30] contains a typo: 'ane et al.' should be 'and et al.' or the correct author list; please check all references for consistency.","section":"References"},{"comment":"The symbol 'offsett' is used as both a word and a variable name; please denote the offset at timestep t by a consistent symbol, such as o_t, and use it consistently in Eqs. (3)-(8).","section":"Section II-B"},{"comment":"The implementation details do not state the number of DDIM steps, the classifier-free guidance scale, or whether the same random seed was used across all methods; these choices affect relative performance and should be reported.","section":"Section III-C"},{"comment":"The claim about 'highest scores in dynamic degree, aesthetic quality, and overall video-text consistency' is faithful to Table I, but the text should also note that FreeNoise attains the highest subject consistency (95.09) and FIFO-Diffusion is higher than the proposed method on subject consistency as well; this nuance is important for a balanced reading.","section":"Section III-D"},{"comment":"The padding of the initial latent to F+2f frames and the cropping of the middle F frames of the output is described in one sentence; please elaborate on how this interacts with the offset schedule and how the first/last short segments are handled in detail.","section":"Section II-B"},{"comment":"The word 'parallized' appears in the contribution list; please correct to 'parallelized'.","section":"Section I"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about Eq. (6) lands: the paper's central mechanism is plausible but unvalidated as a distributional approximation, and the reported metrics are too coarse to catch the specific artifacts the mechanism might introduce. The paper is early-stage and the writing is rough. I would support a major revision rather than rejection, provided the authors add multi-seed evaluation, fix the stride-selection protocol, and provide boundary-focused diagnostics. The idea is distinct from the authors' prior work as far as this manuscript shows."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: the brick-to-wall schedule—denoise f-frame segments, shift by one frame, re-slice—is a distinct and simple mechanism, and the paper describes it cleanly. It is not the same as Gen-L-Video's averaging, FreeNoise's attention fusion, or FIFO's diagonal denoising, and the equations for offset and padding are straightforward. The ablation in Table II is the most convincing piece: s=0 collapses to concatenation and scores drop sharply, so the shifting is doing real work.\n\nWhere the paper is thin is evidence for the headline claim. The main comparison to FIFO-Diffusion rests on a single run, with no error bars or multiple seeds, and the stride s=1 was chosen from the same VBench prompt set used to produce the final table. That is mild test-set tuning, not fraud, but it weakens the 2.6-point average margin, which is driven almost entirely by the dynamic-degree score. There is also exactly one qualitative example, no released code or videos, and the 'arbitrary length' claim is only demonstrated at 128 frames.\n\nThe stress-test worry about Eq. (6) deserves a mention. It's true that independent segment denoising creates artificial boundaries, and shifting the grid only redistributes those boundaries rather than erasing the histories. But the method is explicitly a heuristic, and the ablation shows the stride helps a lot. The real gap is that the paper never measures boundary artifacts or shows convergence toward joint denoising; VBench's aggregate scores are not sensitive to localized flicker. That is an open question, not a demonstrated flaw.\n\nOverall: the idea is worth taking seriously, and the paper is honest enough about its design. It deserves peer review, but the referee should ask for error bars, a held-out stride choice, more qualitative examples, and ideally code. If those come back solid, this is a useful contribution to the training-free long-video subfield.","headline":"A simple, clearly explained training-free trick for long video generation that is plausibly useful, but the current evidence does not yet support the headline margin over FIFO-Diffusion.","tokens_in":8004,"tokens_out":2516,"would_cite":true,"duration_ms":25993,"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":"Brick-Diffusion claims that shifting which frames are denoised together at each step lets a pre-trained short-video diffusion model produce long, consistent videos with no training.","keywords":["diffusion models","long video generation","training-free method","brick-to-wall denoising","temporal consistency","text-to-video generation","latent denoising","inference-time adaptation"],"falsifier":"Generate a 128-frame video of a static scene with a known fixed background and compare it with the same scene generated by the native 16-frame model; if the brick-to-wall video shows periodic flicker or content drift aligned with the 16-frame segment boundaries, the segment-wise approximation is not repairing boundary effects. A more direct check is to denoise the same overlapping frames under two different segment alignments at the same timestep and measure the latent difference, since large differences would mean the model's predictions are inconsistent across segmentations.","tokens_in":7031,"feed_emoji":"🧱","tokens_out":4674,"duration_ms":40800,"temperature":0.7,"pith_summary":"The paper claims that a pre-trained short-video diffusion model can generate arbitrarily long, consistent videos at inference time by denoising the long latent in short segments and shifting the segment boundaries by one frame at each denoising step. This brick-to-wall strategy lets each frame be denoised alongside different neighbours over the course of sampling, so information travels across the whole clip without any training or fine-tuning. The authors report that this method outperforms existing training-free baselines in dynamic degree, aesthetic quality, and video-text consistency on VBench, with the best average score of 81.70 on 128-frame videos. If true, it would remove the need for costly long-video training data and let existing short-video models be reused directly.","feed_headline":"Staggered denoising bricks build long videos without retraining","feed_subtitle":"By shifting which frames are denoised together each step, a pretrained 16-frame model outputs 128-frame clips.","key_machinery":"The brick-to-wall denoising procedure: at each denoising step the long latent is sliced into segments of f frames (f = 16, the pre-trained model's native length), each segment is denoised independently, then the entire latent is shifted by a stride (set to 1 in the main experiments) and re-sliced for the next step. The offset at step t is offset_t = stride × (T − t) mod f, which cycles across the frame dimension, so a frame that was near the edge of one segment becomes interior in a later segment. A shorter first or last segment is handled by temporarily extending it to f frames, denoising, and keeping only the valid part. This shifting is what converts independent segment denoising from plain concatenation into a global communication mechanism.","core_discovery":"On the paper's own terms, the central discovery is that shifting the segmentation of a long latent by a fixed stride between successive denoising steps — instead of denoising fixed, disjoint slices — is enough to keep a long video temporally consistent and visually sharp. Each f-frame segment is denoised independently by the pre-trained model for the current noise level, then the latent is re-sliced with a one-frame offset and denoised again; over many steps every frame is repeatedly co-denoised with many different sets of neighbours, which the paper likens to a staggered brick wall. This simple re-slicing communicates information between segments and avoids both the blurring caused by averaging overlapping windows and the training-inference gap of diagonal denoising. The paper reports that the method achieves the highest dynamic degree, aesthetic quality, and overall text-video consistency among the training-free baselines tested, while FreeNoise retains an edge in subject consistency.","pith_inferences":["The same shifting-segment trick could be applied to other iterative generative models with a fixed context length; the stride acts as a cheap way to couple overlapping contexts.","Combining brick-to-wall denoising with noise rescheduling of the kind FreeNoise uses might improve subject consistency further, since the paper shows FreeNoise's advantage lies there while brick-to-wall leads in dynamics and fidelity.","A direct test of the approximation would be to compare co-denoised overlapping regions: at a given timestep, a frame denoised as part of segment A and the same frame denoised as part of segment B should agree; the paper does not report such an error measure, and this would pinpoint where the joint-denoiser approximation is weakest.","Because the offset formula depends on the total timestep T, the sequence of segmentations is deterministic; a random or adaptive stride schedule might smooth out any residual periodic artifacts every f frames."],"forward_implications":["A pre-trained short-video diffusion model can produce videos of arbitrary length with no fine-tuning, as long as the inference loop re-slices the latent with a non-zero stride.","The stride can be varied (values 1, 3, 5, 7, and 9 all score comparably in the ablation), so the method is robust to the choice of shift.","Because each segment is denoised independently, the procedure can be run in parallel, unlike sequential frame-by-frame or full temporal-attention methods.","The main scores improve with a non-zero stride: average 81.70 at stride 1 versus 72.20 at stride 0 (concatenation), showing the shift is the active ingredient.","The short first and last segments are handled by padding, so the method supports any output length, not only multiples of f."],"supporting_citations":[{"why":"Supplies the pre-trained 16-frame text-to-video model (VideoCrafter2) to which brick-to-wall denoising is applied.","marker":"[10]"},{"why":"DDIM sampler used for every segment denoising step.","marker":"[32]"},{"why":"VBench benchmark and normalized metrics used for quantitative comparison across methods.","marker":"[36]"},{"why":"Gen-L-Video baseline, whose sliding-window averaging is contrasted with brick-to-wall denoising.","marker":"[29]"},{"why":"FreeNoise baseline, whose window-based attention fusion and lack of motion dynamics the method addresses.","marker":"[30]"},{"why":"FIFO-Diffusion baseline, whose diagonal denoising training-inference gap the method avoids.","marker":"[31]"},{"why":"Establishes the diffusion denoising formulation that Equation (6) builds on.","marker":"[2]"}],"fun_headline_variants":["Brick-to-wall denoising: long videos, no retraining","Staggered denoising segments turn short model into long videos","No training, just re-slicing: long videos with brick-to-wall","Brick-Diffusion: 16 frames become 128, training-free","Staggered denoising: brick-by-brick long video generation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that denoising f-frame segments independently at each timestep, with only a shifted regrouping between steps, approximates the joint denoising the pre-trained model would perform on the whole clip; if segment-wise processing shifts the latent distribution or leaves boundary artifacts that the shifting cannot repair, the quality claim fails.","fun_headline_variants_meta":{"raw":{"variants":["Brick-to-wall denoising: long videos, no retraining","Staggered denoising segments turn short model into long videos","No training, just re-slicing: long videos with brick-to-wall","Brick-Diffusion: 16 frames become 128, training-free","Staggered denoising: brick-by-brick long video generation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001323,"raw_usage":{"total_tokens":5364,"prompt_tokens":901,"completion_tokens":4463,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":4366}},"tokens_in":517,"tokens_out":4463,"duration_ms":32340,"temperature":1.0,"reasoning_tokens":4366,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:05:11.899131+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a 128-frame video of a static scene with a known fixed background and compare it with the same scene generated by the native 16-frame model; if the brick-to-wall video shows periodic flicker or content drift aligned with the 16-frame segment boundaries, the segment-wise approximation is not repairing boundary effects. A more direct check is to denoise the same overlapping frames under two different segment alignments at the same timestep and measure the latent difference, since large differences would mean the model's predictions are inconsistent across segmentations.","supporting_citations":[{"cited_title":"Videocrafter2: overcoming data limitations for high-quality video diffusion models,","cited_arxiv_id":null,"evidence_quote":"Supplies the pre-trained 16-frame text-to-video model (VideoCrafter2) to which brick-to-wall denoising is applied."},{"cited_title":"VBench: comprehensive benchmark suite for video generative models,","cited_arxiv_id":null,"evidence_quote":"VBench benchmark and normalized metrics used for quantitative comparison across methods."},{"cited_title":"FreeNoise: tuning-free longer video diffusion via noise rescheduling,","cited_arxiv_id":null,"evidence_quote":"FreeNoise baseline, whose window-based attention fusion and lack of motion dynamics the method addresses."}],"review_version":1}