{"id":"fd261b44-a600-4730-85fa-9ebb271abde2","arxiv_id":"2411.18668","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A k-step search over initial noises reduces quality drift in autoregressive chunk-by-chunk video generation, especially for small image-to-video models.","lead":"This paper tests a cheap way to make long videos from short image-to-video diffusion models: generate the video in chunks, try a few random starting noises for each chunk, and pick the best one using only a few denoising steps. If it works, long video generation becomes practical on smaller models without retraining or huge memory, though the gains shrink on large models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim is not yet validated: Figure 6 shows k-step outputs become closer to full-step outputs, not that Eq-4 scores at k=8 rank noises like full-denoised quality, and the CLIP-min-to-guide metric may select static chunks.","rationale":"The paper's central claim is that k-step evaluation can identify good initial noises for chunk-by-chunk generation, mitigating cumulative degradation. The strongest evidence, Tables 2 and 3, shows VBench improvements for small models, but the mechanism behind those improvements is not established. The reader's weakest-assumption analysis identifies the same core gap: the paper does not show that k-step outputs rank noises like full denoising, and Eq 4 may favor static or guide-similar outputs. I agree with that assessment and consider it the most load-bearing concern. If the k-step ranking premise fails, the method's gains could come from the extra compute spent on multiple candidates rather than from the proposed proxy; if the metric is static-biased, the reported VBench gains may not correspond to genuinely better long videos. Both possibilities are testable with a rank-correlation experiment and a motion-sensitive metric. The paper is otherwise honest: it is training-free, clearly described, and includes an explicit limitations section acknowledging random sampling and no infinite-length guarantee. Those limitations do not resolve the concern, but they also do not make the work unsound; they just leave the central claim under-supported. Therefore the appropriate verdict remains conditional, matching the reader's recommendation, with the requested validation added as a condition.","tokens_in":12199,"tokens_out":4629,"duration_ms":46006,"concrete_test":"Directly validate the selection mechanism: for the same models and prompts as Table 2, sample m=10 noises per guide image, run both k=8 and full 50-step denoising for each noise, then compute (i) the Spearman rank correlation between Eq-4 scores on the k-step outputs and full-step VBench quality scores (averaged over the five reported metrics), and (ii) the Spearman rank correlation between Eq-4 scores on the full-step outputs and the same VBench scores. If either correlation is weak (e.g., ρ < 0.3) across guides, Algorithm 1's selection is not justified. Also compare average optical flow or frame-to-frame difference between k-step-selected and naive chunks; if selected videos have significantly lower motion, the metric is selecting staticness rather than quality.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1's benefit depends on two unvalidated premises. First, Section 3.4 asserts that k-step evaluation 'is able to indicate the overall quality of an otherwise fully denoised output,' but Figure 6 only plots cosine similarity between k-step and full-step videos, not whether candidate noises are ranked the same way. Without a rank-correlation check, the selection step could be choosing near-arbitrarily among m candidates; Tables 2 and 3 would then only show that spending m×k extra denoising steps per chunk improves metrics, not that the proposed proxy works. Second, the selection metric (Eq 4) is the minimum CLIP cosine similarity between each k-step frame and the guide image. This rewards chunks that stay close to the guide; static or near-static chunks trivially maximize this score and also score well on VBench's Subject Consistency, Background Consistency, Temporal Flickering, and Motion Smoothness (AMT reconstructs static frames perfectly). The paper reports no motion-dynamics metric or human evaluation, so the observed VBench improvements may reflect a static bias rather than mitigation of cumulative degradation. Both premises are load-bearing for the central claim, and neither is directly tested in the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses long video generation by autoregressive chunk-by-chunk inference with pretrained image-to-video (I2V) diffusion models. The authors observe that the initial noise strongly influences per-chunk quality and that bad noises cause cumulative degradation. They propose a k-step search: for each chunk, m candidate noises are partially denoised for k steps, scored by the minimum CLIP cosine similarity between the frames and the guide image (Eq. 4), and the winner is fully denoised. Experiments on StableVideoDiffusion, ConsistI2V, OpenSoraPlanV1.3.0, and CogVideoX report VBench improvements for the smaller models. The paper also provides an analysis of noise sensitivity and a comparison of naive versus k-step chunk-wise generation.","tokens_in":12453,"tokens_out":3155,"duration_ms":28443,"significance":"If the central claim is validated, the method is practically useful: it is training-free, compatible with existing I2V models, and could reduce cumulative error accumulation in chunk-wise long-video generation, especially for lightweight models. The paper also contributes a useful empirical analysis of initial-noise sensitivity across four models. However, the validation is currently incomplete: the key proxy assumption (k-step scores rank noises like full denoising) is not directly tested, and the selection metric may induce a static-frame bias that inflates the reported VBench gains. The evidence base is also thin (m and prompt counts unspecified, no variance or significance in the main tables), so the favorable results should be treated as preliminary.","major_comments":[{"comment":"The load-bearing assumption that 'given the same initial noise, the proposed k-step evaluation is able to indicate the overall quality of an otherwise fully denoised output' is not directly tested. Figure 6 only plots cosine similarity between k-step and 50-step outputs as k increases; it does not show that the ordering of candidate noises by Eq. (4) at k=8 matches the ordering by full-denoised quality. Without a rank-correlation or selection-accuracy analysis, the improvements in Tables 2 and 3 could be attributed to the extra computation (m times k additional denoising steps per chunk) rather than to the proposed proxy. Please report the agreement between k-step selected noises and best full-denoised noises, e.g., by comparing Eq. (4) scores at k=8 with full-denoised VBench scores or human ratings over a set of candidates.","section":"Section 3.4, Figure 6"},{"comment":"The selection metric, minimum CLIP cosine similarity to the guide image, may favor static or near-static chunks, because a chunk that stays close to the guide trivially maximizes this score. The reported VBench metrics (Subject Consistency, Background Consistency, Temporal Flickering, Motion Smoothness) also reward static content: AMT reconstructs static frames perfectly, and pixel-level flickering is minimized. The paper reports no motion-dynamics metric or human evaluation, so the observed improvements in Table 2 may reflect a static bias rather than mitigation of cumulative degradation. Please add a quantitative motion metric (e.g., average optical flow magnitude, frame-to-frame difference, or a video motion score) and/or a human preference study, and report qualitative examples showing that motion is preserved.","section":"Section 3.5, Eq. (4)"},{"comment":"Several experimental details needed for reproducibility and significance are missing. The number of candidate noises m used in Algorithm 1 is never specified, and the number of prompts per VBench category is not stated. Tables 2 and 3 report no variance or significance information, unlike Table 1 which reports min/max/std; therefore it is unclear whether the reported improvements are statistically reliable. Please specify m, the prompt count, and report standard deviations or confidence intervals across prompt instances, especially for the small-model improvements that support the central claim.","section":"Section 4.1, Tables 2 and 3"}],"minor_comments":[{"comment":"There is a typo: 'EulerDiscreate' should be 'EulerDiscrete'.","section":"Section 4.1"},{"comment":"The sentence 'we set manual seed before every k-step and full-step sampling' should read 'we set a manual seed'.","section":"Section 4.1"},{"comment":"The variable names 'vs' and 'vcand' are not defined in the pseudocode; consider renaming them to 'output_videos' and 'candidate_videos' for clarity.","section":"Algorithm 1"},{"comment":"The equation numbering is inconsistent: the text says 'Eq. 4' but the equation is not numbered in the text; please add equation numbers consistently.","section":"Section 3.5, Eq. (4)"},{"comment":"The choice k=8 is justified only by a qualitative observation from Figure 6; consider reporting quantitative results, such as selection accuracy at different k values, to support this hyperparameter choice.","section":"Section 4.4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is more of an empirical study and application note than a new method development; the proposed k-step search is simple and the main value is in the analysis and the practical recipe. The editor may wish to consider whether the level of validation is sufficient for the target venue; the missing rank-correlation check and static-bias concern are the key risks."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new piece here is the k-step noise search: sample m noises per chunk, denoise each only k steps, score the half-baked videos, and fully denoise the winner. That is a sensible, training-free extension of best-of-N sampling to autoregressive I2V generation, and the paper is honest about what it does not solve. The authors clearly state they cannot claim infinite-length generation and that their method still relies on random sampling rather than explicit noise refinement. For a short empirical paper, that is the right scope.\n\nWhat the paper does well is motivate the problem. Table 1 concretely shows that initial noise causes large quality variance in small I2V models, and the degradation pattern in chunk-by-chunk generation is real and visible. The VBench improvements for StableVideoDiffusion and ConsistI2V are consistent across four metrics and across chunk counts up to 20. That is credible evidence that the search helps those models. The negative result for large DiT models is also useful and matches the intuition that they are more noise-robust.\n\nNow the soft spots, in proportion. The biggest one is that the core ranking assumption is not directly tested. Section 3.4 asserts that k-step evaluation indicates full-denoise quality, but Figure 6 only shows that k-step outputs become visually closer to full-step outputs as k grows. It does not show that Eq. 4 at k=8 ranks candidate noises the same way full denoising would. Without a rank-correlation or selection-agreement check, the improvement in Tables 2 and 3 could come partly from the extra compute spent on m candidates rather than from the proxy actually identifying good noises. This is a fixable gap: the authors could run full denoising on the same m candidates for a few prompts and report how often the k-step pick matches the full-step pick.\n\nSecond, the selection metric itself deserves scrutiny. Eq. 4 rewards frames that stay close to the guide image in CLIP space, which overlaps with VBench's Background Consistency and could bias toward static or near-static chunks. The paper reports no motion-dynamics metric and no human evaluation, so a static-bias story is not ruled out. I would not call this a fatal flaw, because the authors include Motion Smoothness via AMT, but static chunks also score well there. A simple motion-magnitude statistic or a user study would address it.\n\nThird, the experimental reporting is thin: m is never specified, the number of test prompts is not stated, and Tables 2 and 3 have no variance or significance. These are mundane but necessary details. FreeInit and FrameInit are discussed but not run as baselines, which would have made the cost-quality comparison much stronger.\n\nWho is this for? Anyone working on training-free long-video generation with small I2V models. It deserves a serious referee: the idea is simple, useful, and honestly described, but the evidence needs the missing controls and a direct validation of the ranking assumption before I would rely on the claimed mechanism.","headline":"A practical, honestly-scoped k-step noise search for chunk-by-chunk long video generation, held back by under-reported experiment details and an unvalidated ranking assumption.","tokens_in":12971,"tokens_out":1066,"would_cite":true,"duration_ms":11923,"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 selecting each chunk's initial noise with a cheap k-step denoising probe mitigates the compounding quality loss in chunk-by-chunk long video generation.","keywords":["long video generation","diffusion models","image-to-video","autoregressive generation","chunk-by-chunk generation","initial noise selection","k-step search","training-free inference"],"falsifier":"Run the k-step search with, say, $m=10$ candidate noises on a set of 100 guide images and compare, per image, the k-step ranking against the ranking of the same noises after full denoising measured by VBench subject consistency. If the k-step argmax matches the full-step argmax at no better than chance, or if the k-step winner's VBench score is not reliably above a randomly chosen noise, the central claim collapses.","tokens_in":11951,"feed_emoji":"🎬","tokens_out":4955,"duration_ms":42173,"temperature":0.7,"pith_summary":"Large diffusion video models are typically trained on fixed, short clips, so generating a long video in one pass exceeds GPU memory and drifts out of distribution. This paper studies a memory-feasible alternative: generate the video chunk by chunk, with an image-to-video model that each time animates the last frame of the previous chunk. The paper's central claim is that in this autoregressive setup the initial noise sampled for each chunk strongly determines whether that chunk stays clean, and that bad noises cause errors to compound across chunks. To exploit this, it proposes a k-step search that denoises several candidate noises for only k steps, scores the resulting rough videos by their minimum CLIP cosine similarity to the guide frame, and fully denoises only the best candidate. The experiments show the search meaningfully reduces degradation for smaller UNet-based I2V models, while the largest DiT models are stable enough that naive chunk-by-chunk generation already works.","feed_headline":"Cheap noise probe keeps long videos from degrading","feed_subtitle":"A k-step search picks the best initial noise for each chunk, letting small image-to-video models stay clean across many chunks.","key_machinery":"The central mechanism is the k-step search (Algorithm 1) combined with the train-inference noise discrepancy that motivates it. During training a diffusion model sees noisy latents of the form $z_t = \\sqrt{\\bar\\alpha_t} z_0 + \\sqrt{1-\\bar\\alpha_t}\\epsilon$, while at inference it is handed a pure Gaussian $z_T \\sim \\mathcal{N}(0,I)$; these distributions are close but not identical, so some initial noises lie outside the region the model learned to denoise well. The k-step search compensates by trying $m$ noises, denoising each for $k$ steps, and scoring the rough result with $\\min_j \\mathrm{CosSim}(\\mathrm{CLIP}(I_{\\text{guide}}, V[j]))$ over frames $j$. The low-step probe is cheap, and the chosen noise then runs the full denoising schedule.","core_discovery":"The paper establishes that, for a given image-to-video diffusion model and conditioning frame, different Gaussian initial noises lead to markedly different chunk quality (Table 1), and that the worst of these noises is what drives the cumulative worsening effect in chunk-by-chunk generation. Its proposed cure is the k-step search (Algorithm 1): sample m noises, run the denoiser for k steps on each, score the suboptimal videos with the minimum CLIP cosine similarity between their frames and the guide image, then fully denoise the noise that scores highest. The discovery is that such a short probe, with k as small as 8, is enough to rank noises well enough to select a good one, so the per-chunk error never gets a foothold. This is demonstrated on four I2V models, with the largest gains on the smaller UNet-based models and only marginal gains on large DiT models.","pith_inferences":["The ranking assumption between k-step and full-step quality is directly testable: on any I2V model, one can compute the fraction of times the k-step argmax matches the full-step argmax across many guide images; if that fraction is near chance for small k, the probe would need more steps.","The minimum-CLIP-similarity score may implicitly penalize motion and favor static or semantically close frames, so a temporal-aware selection metric could change which noise wins and further improve motion quality.","Because the paper notes error never fully disappears, a closed-loop monitor that re-runs the k-step probe every few chunks and resamples when the score drops could push generation beyond the demonstrated 20 chunks.","k-step search could be combined with noise-refinement methods that alter low-frequency components, turning pure random search into a two-stage candidate generation that might outperform either alone."],"forward_implications":["Chunk-by-chunk generation with an off-the-shelf I2V model can produce long videos without out-of-memory failures, and the k-step search keeps quality degradation in check for smaller models.","Small UNet-based I2V models, which are fast but noise-sensitive, benefit most from k-step selection; large DiT models need little intervention.","The overhead of k-step search is modest: with $m=10$ candidates and $k=8$ steps, the extra cost is 80 steps per chunk compared with 500 for brute force over 10 candidates when full sampling is 50 steps.","k-step search is training-free and works with any I2V model whose scheduler allows a reduced step count, as long as the initial noise can be fixed by a random seed."],"supporting_citations":[{"why":"Supplies the denoising diffusion formulation and the predefined noise schedule that the whole method operates on.","marker":"[7]"},{"why":"Provides the DDIM scheduler that allows reduced-step sampling and reproducible outputs given the same initial noise, enabling the k-step probe.","marker":"[23]"},{"why":"Defines CLIP, whose embedding cosine similarity is the selection score in Eq. (4).","marker":"[18]"},{"why":"Stable Video Diffusion is one of the small UNet-based I2V models used to demonstrate the degradation problem and the k-step improvement.","marker":"[1]"},{"why":"ConsistI2V is another small I2V testbed and comparison baseline for chunk-by-chunk generation.","marker":"[20]"},{"why":"OpenSoraPlan V1.3.0 is the large DiT model that shows naive chunk-by-chunk generation already works well.","marker":"[12]"},{"why":"CogVideoX is the other large DiT I2V model used to show the k-step search gives only marginal benefit.","marker":"[28]"},{"why":"VBench provides the evaluation metrics for subject consistency, background consistency, temporal flickering, motion smoothness, and aesthetic quality.","marker":"[8]"}],"fun_headline_variants":["k-step noise search stops long video worsening","Probe noise briefly to keep chunked videos clean","Pick better noise per chunk for long videos","k-step search chooses noise to avoid video decay","Short noise probe stops long video degradation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"That a video denoised for only k steps ranks candidate initial noises in the same order of quality as a fully denoised video, and that the minimum CLIP similarity of those k-step frames to the guide image is a trustworthy proxy for that quality.","fun_headline_variants_meta":{"raw":{"variants":["k-step noise search stops long video worsening","Probe noise briefly to keep chunked videos clean","Pick better noise per chunk for long videos","k-step search chooses noise to avoid video decay","Short noise probe stops long video degradation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000531,"raw_usage":{"total_tokens":2559,"prompt_tokens":946,"completion_tokens":1613,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":1545}},"tokens_in":562,"tokens_out":1613,"duration_ms":13178,"temperature":1.0,"reasoning_tokens":1545,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:08:57.070382+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the k-step search with, say, $m=10$ candidate noises on a set of 100 guide images and compare, per image, the k-step ranking against the ranking of the same noises after full denoising measured by VBench subject consistency. If the k-step argmax matches the full-step argmax at no better than chance, or if the k-step winner's VBench score is not reliably above a randomly chosen noise, the central claim collapses.","supporting_citations":[{"cited_title":"Denoising dif- fusion probabilistic models","cited_arxiv_id":null,"evidence_quote":"Supplies the denoising diffusion formulation and the predefined noise schedule that the whole method operates on."},{"cited_title":"Open-sora-plan, 2024","cited_arxiv_id":null,"evidence_quote":"OpenSoraPlan V1.3.0 is the large DiT model that shows naive chunk-by-chunk generation already works well."}],"review_version":1}