{"id":"7bd79cd6-c175-45f3-9373-0d60f06772bc","arxiv_id":"2502.05433","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"AdaFlow demonstrates a training-free method to edit more than 1,000 video frames in one inference on a single A800 GPU via adaptive attention token slimming and content-aware keyframe selection.","lead":"AdaFlow edits very long videos, over 1,000 frames, in a single pass on one GPU by trimming the tokens used in attention during keyframe translation and by picking keyframes based on video content. It also introduces a new benchmark of one-minute videos for evaluating long-video editing.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim rests on the untested assumption that DIFT-similarity-based KV pruning (Eq. 7) preserves the attention information needed across all 50 denoising steps; no AAS ablation or retained-token sensitivity analysis is provided.","rationale":"The reader identified the same load-bearing point: the DIFT-similarity token selection in Eq. (7) is the key mechanism enabling the order-of-magnitude increase in keyframes, and its sufficiency is not independently validated. I agree and sharpen the concern: the selection is fixed from clean-frame DIFT at t=0 and reused at every denoising timestep, while the attention needs of noisy latents evolve over the 50-step schedule. The paper's own evidence is missing the critical AAS ablation — Figure 3(b) ablates only AKS, Table 1 compares complete systems, and no sensitivity analysis over m is given. This is not an accusation of fraud or a claim that the method is wrong; the released code and qualitative results are real evidence in the paper's favor. But the central claim that AAS is both efficient and quality-preserving is currently supported only by indirect, whole-system comparisons. The CONDITIONAL verdict already captures this: the paper should add an AAS ablation and a token-budget sweep, or explicitly state the quality/efficiency trade-off. Therefore I do not move the verdict; I would keep it CONDITIONAL and require the AAS validation as a condition.","tokens_in":16054,"tokens_out":7050,"duration_ms":70698,"concrete_test":"Using the released code, instrument the UNet self-attention at an intermediate denoising step (e.g., t=25) for a subset of LongV-EVAL videos and a keyframe count that fits in memory; compute the fraction of softmax attention mass assigned to the KV tokens that AAS discards. If this fraction is non-negligible (>5%) averaged over queries and steps, the assumption behind Eq. (7) fails and AAS cannot be considered quality-preserving. As a secondary check, sweep the retained-token budget m and confirm that the default (14 frames' worth) sits on a quality plateau; if FQ/VQ/OC/SC degrade monotonically with m, the paper should report this trade-off explicitly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"AdaFlow's headline capability ('>1k frames in one inference via Adaptive Attention Slimming') rests entirely on the validity of Eq. (7): K and V are replaced by the m tokens whose source-frame DIFT cosine similarity to the query keyframe is highest. For this to preserve editing quality and temporal consistency, those m tokens must carry the attention mass that matters at every denoising step. The selection is computed once from clean source frames (DIFT at t=0; see Section 4.2 and Section 5.2) and reused for all 50 timesteps, but the latents being denoised change substantially over the schedule, so the top-m set at t=0 need not be the top-m set at later timesteps. Moreover, the paper reports no ablation of AAS: Table 1 compares AdaFlow against other methods, Figure 3(b) ablates only AKS, and there is no sweep over the retained-token budget m nor a comparison against full-KV attention at matched keyframe count. Thus the paper does not demonstrate that AAS is quality-preserving; it only shows that the full system scores well on metrics that heavily reward staying close to the source. The Appendix F admission that shape-changing edits often fail is consistent with the hypothesis that the DIFT proxy discards tokens required for object-level consistency. This is the load-bearing risk to the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AdaFlow, a training-free method for long text-driven video editing. The approach divides the video into content-based segments via Adaptive Keyframe Selection (AKS), which uses DIFT features to measure frame similarity and chooses representative keyframes per segment. The keyframes are jointly edited using extended self-attention, but to fit more keyframes in memory, Adaptive Attention Slimming (AAS) prunes the KV token sequence: for each query keyframe, only the m pixel positions with highest DIFT cosine similarity to the query are retained across all keyframes (Eq. 7). The edited keyframe latents are then propagated to non-keyframes using precomputed DIFT correspondences. The paper also introduces a new benchmark, LongV-EVAL, consisting of 75 one-minute videos with three annotated editing prompts each, and reports experiments on an A800 GPU showing editing of more than 1,000 frames in a single inference, with better or comparable quality metrics and lower runtime than several baselines. The authors release their code.","tokens_in":16381,"tokens_out":2947,"duration_ms":28823,"significance":"If validated, AdaFlow's headline result — order-of-magnitude longer single-inference video editing through attention slimming — would be practically significant, and LongV-EVAL could fill a gap in long-video editing evaluation. The paper has clear strengths: it is training-free, the pipeline is described in sufficient detail to be reproduced, the code is released, qualitative results include 10k-frame examples (Appendix B, Fig. 6), and the benchmark provides a standardized set of long videos with prompts targeting foreground, background, and style edits. However, the central claim rests on an untested assumption about the KV-pruning proxy, and the quantitative evidence presented does not yet isolate the contribution of the main component, AAS. The user-study inconsistency and missing prompt-fidelity metric further weaken the current evaluation.","major_comments":[{"comment":"The central claim that Adaptive Attention Slimming enables an order-of-magnitude increase in editable frames while preserving quality is not supported by any ablation of AAS itself. Figure 3(b) ablates only Adaptive Keyframe Selection, and Table 1 compares the full AdaFlow system against baselines. The paper needs a comparison of AdaFlow with and without AAS at matched keyframe count and video length, plus a sweep over the retained-token budget (the '14-frame token count' setting in §5.2), to show that the pruned attention preserves editing quality relative to full-KV extended self-attention. Without this, the headline capability cannot be attributed to AAS, and the limitation admitted in Appendix F (unsatisfactory results when object shapes change) is consistent with the hypothesis that the DIFT-similarity proxy discards tokens required for object-level consistency.","section":"§4.2, Eq. (7); §5.3, Table 1; Fig. 3(b)"},{"comment":"The DIFT-based token selection used by AAS is computed from clean source frames at t=0 and then reused for all 50 DDIM denoising steps, because the latents being denoised change substantially over the schedule. The paper provides no evidence that the set of m most-similar tokens at t=0 remains the set of tokens that carry the attention mass needed at later timesteps. This is a load-bearing assumption for Eq. (7); a sensitivity analysis comparing the fixed t=0 selection against per-timestep selection, or at least reporting the overlap of selected tokens across timesteps, would test whether this assumption holds.","section":"§4.2 and §5.2 (DIFT extraction at t=0)"},{"comment":"The quantitative evaluation lacks error bars, statistical significance tests, and a direct prompt-fidelity metric. The four reported metrics (FQ, VQ, OC, SC) measure aesthetic/technical quality and temporal/object consistency, but none measures whether the output actually follows the editing prompt (e.g., CLIP text-image similarity against the prompt). The differences in OC and SC between AdaFlow and the best baseline are small (0.953 vs 0.947 and 0.969 vs 0.966), and without variance or significance testing the claim in §5.3 of 'obvious advantages' over the compared methods is not established. The user study also has an internal inconsistency: Table 2 and the main text state 18 participants, while Appendix C states 50 participants. This discrepancy must be resolved, and the study protocol (20 sets of 6 videos, 40 choices per participant) should be clarified so that the reported percentages are internally consistent.","section":"§5.3, Table 1 and §5.3 user study; Appendix C"}],"minor_comments":[{"comment":"The 'window check' function in Algorithm 1 is not defined in the main text; the parameters l, s, ws are given in §5.2 but the precise procedure for the sliding-window similarity test should be specified, since it is a core part of AKS.","section":"§4.1, Algorithm 1"},{"comment":"The phrase 'We consistently retain the token count corresponding to 14 frames' is ambiguous: clarify whether this is 14 × h × w tokens per attention layer, and how the m positions are aggregated across keyframes (e.g., per query keyframe independently, or as a union).","section":"§5.2, keyframe pruning setting"},{"comment":"The benchmark name 'LongV-EV AL' appears with a stray space in several places (e.g., §5.1, Table 1); this should be corrected to 'LongV-EVAL' consistently.","section":"§5.1 and Table 1"},{"comment":"The runtime comparison in Table 1 is confounded by the fact that baselines are edited in segments of 128, 32, or 16 frames, as stated in §5.2; reporting per-frame or per-keyframe timing and GPU memory would make the efficiency claim more interpretable.","section":"§5.2, Table 1, timing comparison"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is plausible and the benchmark contribution is useful, but the central efficiency-quality claim currently lacks a decisive experiment: an AAS on/off ablation at matched settings. The user-study participant inconsistency (18 vs 50) is a factual error that the authors must correct, and the absence of any prompt-fidelity metric makes the quality comparison hard to interpret. I would not reject the paper, as the issues are fixable with additional experiments and clarifications within the manuscript's scope, but they are load-bearing for the current claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nThe main thing to know: AdaFlow is a training-free extension of TokenFlow that adds two heuristics — content-based keyframe selection using DIFT similarity, and KV-token pruning in extended self-attention (AAS). The AAS idea is clever, and the qualitative results on 1k–10k frame videos are genuinely impressive. What is new is the combination; neither component alone is novel. AKS is a natural improvement over uniform sampling, and AAS is a plausible way to cut memory. The paper also releases code and a new 75-video benchmark (LongV-EVAL), which is a real asset. The citation pattern is conventional; it builds on the expected set of video editing and DIFT work without over-self-citing.\n\nThe soft spots are in the evaluation. The biggest one: there is no ablation of AAS. Table 1 compares the full system against baselines, and Figure 3 only ablates AKS. Without removing AAS or sweeping the retained-token budget m, we cannot tell whether the pruning preserves quality or whether the gains come purely from keyframe selection and propagation. The stress-test concern about Eq. (7) is valid: the token selection is computed once from DIFT at t=0 and reused for all 50 denoising steps, but the latents change substantially over the schedule. The paper's own Appendix F admits that shape-changing edits often fail, which is consistent with the worry that the pruned tokens carry object-level information.\n\nThere are smaller issues: no error bars or significance tests, no prompt-fidelity metric, and an inconsistency in the user study size (18 in the main text, 50 in Appendix C). The 'ten times longer' claim is a bit loose, since baselines are run in segments of 128 frames and the comparison is not on a common keyframe budget.\n\nNone of this is fatal; the method is plausible and the problem is important. But the central claim — that AAS enables an order-of-magnitude increase in edit length without quality loss — is not yet demonstrated. A serious referee should ask for an AAS ablation, a sensitivity analysis over m, and ideally some analysis of how stable the selected tokens are across denoising steps.\n\nI'd like to see a revision with those experiments. If they hold up, this becomes a useful paper for anyone working on long video editing. As it stands, the evidence is decent but not conclusive.\n\nRecommendation: send to peer review, but expect major revision.","headline":"Plausible training-free long-video editing with a clever KV-pruning trick, but the paper never isolates that trick in an ablation, leaving the central claim under-supported.","tokens_in":16890,"tokens_out":5001,"would_cite":true,"duration_ms":42617,"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":"AdaFlow claims that text-driven long video editing can be scaled to thousands of frames by trimming extended self-attention to the tokens most similar to each query frame, cutting memory while preserving consistency.","keywords":["long video editing","text-driven video editing","diffusion models","extended self-attention","attention slimming","keyframe selection","DIFT correspondence","training-free editing"],"falsifier":"A concrete test: pick a source video where a subject changes shape during motion, such as a person turning or an animal bending, apply AdaFlow with the same prompt under full extended self-attention and under AAS with progressively smaller m, and measure object consistency and edit fidelity on the shape-changing frames; if consistency collapses as m falls, or if the m retained positions per query do not cover the edited object, the central assumption is falsified. A second check is to run AdaFlow on a long static scene versus a rapid-cut video and compare whether quality degrades with length under a fixed m budget.","tokens_in":15862,"feed_emoji":"🎬","tokens_out":7322,"duration_ms":66274,"temperature":0.7,"pith_summary":"The paper's target is the memory wall of text-driven long video editing: extended self-attention over all keyframes grows quadratically in the number of tokens, so existing methods cap out at a few hundred frames. AdaFlow argues that for each query frame most KV tokens in that attention are irrelevant, and that the relevant ones can be found by DIFT cosine similarity between frames. It therefore trims K and V to the m most similar token positions per query, and selects keyframes by content change rather than uniform sampling. With this design, the paper reports editing videos of more than 1,000 frames in a single inference on one A800 GPU, about ten times longer than TokenFlow, while keeping or improving consistency scores. It also contributes LongV-EVAL, 75 one-minute videos annotated with three editing prompts each, as a long-video benchmark.","feed_headline":"Attention slimming lets one GPU edit 1,000-frame videos","feed_subtitle":"AdaFlow drops most key/value tokens per query, cutting memory so long videos can be edited in one pass without retraining.","key_machinery":"The load-bearing mechanism is Adaptive Attention Slimming (AAS), applied inside the extended self-attention of keyframe translation. For one query keyframe, DIFT cosine similarity heatmaps are built against every other keyframe; only the m pixel positions with the highest similarity survive, and K and V are replaced by eK and eV holding just those tokens, so Eq. (3) becomes Eq. (7). This shrinks the KV sequence length from the full set of keyframe tokens to a fixed number, directly attacking the quadratic memory cost of extended self-attention. The same DIFT heatmaps also drive Adaptive Keyframe Selection (AKS), which segments the video into clips of similar content using mean and sliding-window thresholds in Algorithm 1, and Feature-Matched Latent Propagation, which reuses corresponding source tokens to transfer keyframe outputs to non-keyframes.","core_discovery":"On its own terms, the discovery is that the quadratic cost of extended self-attention in keyframe translation can be cut without sacrificing editing quality by observing that a query frame mainly needs tokens similar to itself. AdaFlow computes DIFT similarity heatmaps between each query keyframe and every other keyframe and retains only the m highest-similarity positions in K and V, replacing them with slimmed eK and eV in Eq. (7). That reduces the KV sequence from M times h times w tokens to a fixed budget, which is what allows the number of keyframes edited together to rise by an order of magnitude. Adaptive Keyframe Selection then distributes the budget according to content dynamics, and Feature-Matched Latent Propagation spreads keyframe edits to the remaining frames using correspondences computed once. The paper claims this yields high-quality editing of more than 1k frames in one inference, with quantitative results on LongV-EVAL and a user study favoring AdaFlow over five baselines.","pith_inferences":["Beyond the paper: the same DIFT-based token selection could be applied to temporal attention in other diffusion video pipelines, not just the keyframe-translation stage, since the underlying quadratic cost is shared.","Beyond the paper: a fixed token budget suggests a testable scaling law, namely that object consistency should degrade with video length under AAS even if per-frame edit quality stays constant; measuring that curve would tell how far the method can scale.","Beyond the paper: the reliance on DIFT similarity makes AdaFlow more suited to appearance and style edits than to geometry-changing edits, and the paper's Appendix F concedes exactly this weakness; a stress test separating the two edit types would locate the practical boundary of the method.","Beyond the paper: a natural extension would be a variant that restores full attention only for tokens near regions named by the edit prompt, combining the memory savings with better shape-change handling."],"forward_implications":["Videos of more than 1,000 frames can be edited in one inference on a single 80GB GPU, and the appendix reports one 10,000-frame edit.","Editing time drops: AdaFlow averages 24 minutes per LongV-EVAL video, half or less of the 40 to 83 minutes reported for the five baselines.","Because AAS is training-free, the method inherits improvements in the underlying image diffusion backbone without retraining.","Keyframe density becomes content-adaptive: static segments consume fewer keyframes, while rapid-motion segments get denser sampling, which avoids both redundancy and blur around motion.","The evaluation protocol of LongV-EVAL, covering frame quality, video quality, object consistency, and semantic consistency, can be reused to compare future long-video editors."],"supporting_citations":[{"why":"Defines the two-step keyframe-translation and interpolation paradigm plus extended self-attention that AdaFlow extends, and serves as the main baseline.","marker":"(Geyer et al., 2023)"},{"why":"Supplies DIFT, the diffusion features whose cosine similarity is the basis for attention slimming, keyframe selection, and propagation.","marker":"(Tang et al., 2023)"},{"why":"Provides the PnP-Diffusion image editing technique used to translate each keyframe.","marker":"(Tumanyan et al., 2023)"},{"why":"Supplies the Stable Diffusion backbone whose latent self-attention layers AdaFlow edits and slims.","marker":"(Rombach et al., 2022)"},{"why":"Provides DDIM inversion, used in preprocessing to produce the latents AdaFlow denoises during editing.","marker":"(Song et al., 2020)"},{"why":"Rerender is a compared two-step baseline that also samples keyframes, and its official implementation is used for comparison.","marker":"(Yang et al., 2023)"},{"why":"DINO embeddings are used to compute the object consistency metric reported on LongV-EVAL.","marker":"(Caron et al., 2021)"},{"why":"CLIP embeddings are used to compute semantic consistency between adjacent edited frames.","marker":"(Radford et al., 2021)"}],"fun_headline_variants":["One GPU edits 1,000 frames with slimmer attention","AdaFlow slims attention to edit minute-long videos","Cut KV tokens to extend video editing to 1k frames","Training-free method edits 10x longer videos on one GPU","Adaptive slimming yields 10x more keyframes for editing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that keeping only the m pixel tokens most similar to the query frame, measured by DIFT cosine similarity, preserves the information needed to produce a consistent edit, so the slimmed attention is as good as full attention; the paper's Appendix F acknowledges this can fail when edits require object-shape changes.","fun_headline_variants_meta":{"raw":{"variants":["One GPU edits 1,000 frames with slimmer attention","AdaFlow slims attention to edit minute-long videos","Cut KV tokens to extend video editing to 1k frames","Training-free method edits 10x longer videos on one GPU","Adaptive slimming yields 10x more keyframes for editing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000328,"raw_usage":{"total_tokens":1846,"prompt_tokens":976,"completion_tokens":870,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":785}},"tokens_in":592,"tokens_out":870,"duration_ms":8564,"temperature":1.0,"reasoning_tokens":785,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T19:20:39.356207+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: pick a source video where a subject changes shape during motion, such as a person turning or an animal bending, apply AdaFlow with the same prompt under full extended self-attention and under AAS with progressively smaller m, and measure object consistency and edit fidelity on the shape-changing frames; if consistency collapses as m falls, or if the m retained positions per query do not cover the edited object, the central assumption is falsified. A second check is to run AdaFlow on a long static scene versus a rapid-cut video and compare whether quality degrades with length under a fixed m budget.","supporting_citations":[],"review_version":1}