{"id":"c8cc52a0-18ac-4c53-b3a5-80c170f39266","arxiv_id":"2505.07057","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A dual-stage fine-tuning recipe, norm tuning followed by a visual adapter, improves temporal consistency and text alignment in one-shot video editing, with a new 232-video benchmark.","lead":"This paper presents DAPE, a two-stage parameter-efficient fine-tuning method that improves video editing by first tuning normalization layers for temporal smoothness, then adding a visual adapter for visual quality. It also introduces a new 232-video benchmark with six editing prompts per video to support systematic comparison of video editing models.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains may stem from per-video test-time adaptation rather than the proposed PEFT modules; Table 1 lacks matched controls, error bars, and multi-seed statistics, so 'significantly outperforms' is not established.","rationale":"The reader's verdict is CONDITIONAL with the weakest assumption being that per-video fitting learns transferable editing behavior rather than memorizing source appearance. I agree that test-time adaptation is the crux, but the more precise and falsifiable concern is that Table 1 compares methods with unequal access to per-video training. The largest improvements appear exactly where the baseline had no per-video adaptation, while the matched one-shot baseline (TAV) shows negligible or negative changes. This makes the 'outperforming previous SOTA' claim fragile regardless of whether the learned weights generalize to unseen videos, because the comparison itself may be unfair. The paper's own ablations (Table 3) show that the proposed modules' gains over the baseline are small in absolute terms, and no significance testing or alternative PEFT control is provided. These issues are fixable in a revision, so the CONDITIONAL verdict remains appropriate; no change to the reader's verdict is needed. The concrete test proposed would settle whether the modules matter beyond generic test-time adaptation and whether the headline comparison is confounded.","tokens_in":20218,"tokens_out":5769,"duration_ms":62112,"concrete_test":"Run a controlled comparison on at least 20 held-out videos with matched compute: (a) DAPE as described; (b) the same two-stage schedule but with a randomly initialized visual adapter and frozen norm parameters; (c) per-video LoRA with a matched number of trainable parameters; (d) per-video fine-tuning of the same norm layers without the adapter. Compute all five metrics from Table 1 over at least three seeds and report 95% confidence intervals. If variant (c) or (d) matches DAPE within noise, the specific norm-tuning/adapter design is not the cause of the reported gains. Also report trainable parameter counts and wall-clock training time per video for each variant, and verify whether the DAPE-versus-baseline gap persists when RAVE and CCEdit are also given per-video adaptation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DAPE's dual-stage norm-plus-adapter fine-tuning significantly improves temporal consistency and text-video alignment over prior SOTA. The main evidence is Table 1, where each DAPE row is obtained by fine-tuning on the same source video that is later edited (Sec. 3.2, Eq. 3-6; Sec. 5.2). This protocol conflates two effects: (i) the specific norm-tuning and visual-adapter design, and (ii) generic test-time adaptation to the target clip. The baselines are not matched in this respect: RAVE is training-free and CCEdit is trained offline, so DAPE(RAVE) and DAPE(CCEdit) add per-video training that the baseline never had. The largest deltas, such as 34.98% War. Err. on the DAPE Dataset for DAPE(CCEdit) and 18.09%/9.82% on the RAVE Dataset, occur precisely for these unmatched baselines, while the matched per-video baseline DAPE(TAV) versus TAV shows near-zero or even negative deltas (e.g., -3.98% Int. Err. on BalanceCC). In addition, no alternative PEFT method (LoRA, trainable norms alone, random adapter) is compared, and the paper provides no error bars, significance tests, or multi-seed runs, so the word 'significantly' is unsupported. The construction of each DAPE variant is also under-specified: Sec. 5.1 gives hyperparameters only for the CCEdit-initialized model and does not state how modules are inserted into TAV, CAMEL, SimDA, RAVE, or CCEdit. The reader's worry about memorization is related, but the more immediate threat is that the headline comparison is confounded by unequal access to test-time adaptation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DAPE, a dual-stage parameter-efficient fine-tuning (PEFT) framework for text-guided video editing. In Stage I, adjustable norm-tuning tunes the affine parameters of normalization layers plus a learnable residual scale to improve temporal consistency; in Stage II, a vision-friendly adapter (down-projection, depth-wise 5x5 convolution, up-projection, with a layer-normalization and skip connection) is inserted at the first cross-attention block of the decoder to improve visual quality. The authors argue that jointly training both modules produces negative interactions, motivating the two-stage schedule. They also introduce the DAPE Dataset, a curated benchmark of 232 videos with category/complexity annotations and multiple edit prompts. Experiments on BalanceCC, LOVEU-TGVE, RAVE, and the DAPE Dataset compare DAPE variants applied on top of five baselines (Tune-A-Video, CAMEL, SimDA, RAVE, CCEdit), reporting CLIP-Frame, interpolation error/PSNR, warping error, and CLIP-Text, plus a user study. The main quantitative evidence is Table 1 with per-baseline deltas.","tokens_in":20616,"tokens_out":7739,"duration_ms":66912,"significance":"If the central claims were established with matched controls, DAPE would be a practically useful contribution: it offers a parameter-efficient way to adapt a video-editing diffusion model to a single clip while decoupling temporal and visual adaptation, and the paper includes useful ablations of adapter placement and module interactions. The dataset contribution is also potentially valuable for standardized evaluation. However, the current evaluation does not establish the headline claim: the protocol trains on the same clip that is later edited, so the gains could be due to generic per-video test-time adaptation rather than the proposed modules, and no error bars or significance tests are provided. The central 'significantly outperforms' claim therefore needs additional experimental controls before it can be accepted. Strengths of the manuscript include the clear architectural description of both stages, the thorough dataset-construction pipeline in the supplementary material, and the honest reporting of the negative interaction that motivates the two-stage design.","major_comments":[{"comment":"The evaluation protocol confounds per-video fine-tuning with the proposed method design. DAPE is fine-tuned on the same source video that is later edited (§3.2), whereas the baselines RAVE (training-free) and CCEdit (offline-trained) are not per-video adapted. The largest deltas (e.g., DAPE(CCEdit) on RAVE Dataset: Int. Err. -18.09%, War. Err. -34.98%) appear exactly where the baseline lacks per-video training. A matched control is required, such as per-video full fine-tuning or per-video LoRA/adapter-only tuning of the same backbone with the same training budget, to separate generic test-time adaptation from DAPE's specific design. Without this, the comparison is not apples-to-apples and the central claim is not established.","section":"§5.2, Table 1"},{"comment":"No error bars, multi-seed runs, or significance tests are reported. Many deltas are below 1% or negative (e.g., DAPE(TAV) on BalanceCC CLIP-T 0.00%, DAPE(CAMEL) on loveu-tgve War. Err. -0.99%, DAPE(SimDA) on loveu-tgve CLIP-T -0.55%), so the word 'significantly' in the abstract is unsupported. Please report variance across at least three seeds and use a paired statistical test or confidence intervals for the main comparisons.","section":"§5.2, Table 1"},{"comment":"The instantiation of DAPE on each baseline is underspecified. Section 5.1 gives hyperparameters only for the CCEdit-initialized model and does not state how adjustable norm-tuning and the visual adapter are inserted into TAV, CAMEL, SimDA, RAVE, or CCEdit. Given the architectural differences (RAVE uses grid concatenation and noise shuffling; CCEdit uses a trident network), the insertion rules are nontrivial and must be specified for reproducibility and for interpreting the per-baseline gains.","section":"§5.1 and Table 1"},{"comment":"Transferability of the fine-tuned modules is not tested. DAPE is tuned on the clip that is later edited, and all metrics are computed on edits of that same clip, so the improvements may reflect memorization of the source appearance rather than generalizable editing behavior. The paper provides no held-out experiment, such as fine-tuning on clip A and editing clip B, or evaluating on prompts unseen during tuning. Please add such a transfer experiment, or explicitly scope the claims to per-video test-time adaptation.","section":"§3.2 / §5.2"},{"comment":"The claim that 'DAPE achieves the best performance across all datasets' is contradicted by several per-metric results in Table 1: DAPE(CCEdit) on BalanceCC has higher Int. Err. than CCEdit (7.58 vs 7.29), DAPE(CCEdit) on loveu-tgve has lower CLIP-F than CCEdit (93.76 vs 94.00), and DAPE(CAMEL) on loveu-tgve has higher War. Err. than CAMEL (4.07 vs 4.03). Please report per-metric wins and losses transparently and avoid the blanket 'outperforming previous state-of-the-art' claim unless an aggregate or statistical comparison justifies it.","section":"§5.2"}],"minor_comments":[{"comment":"The abstract says '6 editing prompts,' while §4 and Figure 9 describe five editing perspectives with two random-combination prompts (six total); align the wording.","section":"Abstract / §4"},{"comment":"The notation 'f = z + ω_dw ⊗_dw z_down' is ambiguous; the initial z should likely be z_norm, and the roles of down-projection and depth-wise convolution should be defined explicitly.","section":"Eq. (5)"},{"comment":"Define 'w/o All' clearly (presumably the RAVE baseline on the RAVE dataset) and specify which dataset the ablation uses.","section":"Table 3"},{"comment":"Clarify whether the pre-trained ControlNet from CCEdit is used during inference for all baselines or only for DAPE variants, since this affects fairness.","section":"§5.1"},{"comment":"Indicate which norm layers are tuned in Stage I (all LayerNorm and GroupNorm blocks?) and which remain frozen.","section":"Figure 2"},{"comment":"The dataset name 'DAPE Dataset' is the same as the method name; consider renaming to avoid confusion in future citations.","section":"§4.1"},{"comment":"User study numbers are inconsistent: 30 participants × 25 sets would be 750 sets, but the text reports 1,536 responses, and the supplementary mentions 21 video-text pairs; clarify the counting.","section":"§5.2 / Appendix B"},{"comment":"Reference [7] appears twice in the bibliography; merge the duplicate entries.","section":"References"},{"comment":"'LOVEU-TGVE' is typeset inconsistently as 'loveu-tgve' in Table 1; standardize the capitalization.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The main issue is the evaluation protocol: per-video fine-tuning on the test clip, combined with the absence of matched per-video baselines and statistical tests, makes the headline claim unsupported in its current form. The self-citation cluster around the adapter/norm motivation (refs 63-65) is concentrated on the corresponding author's prior work; this is not inherently problematic, but independent citations would strengthen the framing. No code or dataset release link is provided; since the contribution includes a benchmark, releasing the dataset with prompts and annotations is important for the field."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before you look at the paper. First, the core idea is genuinely reasonable: fine-tune only norm parameters to stabilize temporal structure, then add a lightweight visual adapter, and sequence those stages because joint training hurts. Second, the headline claim that this 'significantly outperforms' prior SOTA is not supported by the reported experiments, mainly because the method fine-tunes on the same video it then edits, while the baselines it looks best against (RAVE, CCEdit) are not given that test-time adaptation.\n\nWhat is actually new: the dual-stage schedule, the adjustable norm-scaling term gamma_0, and a curated 232-video benchmark with annotations and five editing-prompt types. The ablation in Table 3 does show dual-stage beating one-stage, which is a useful empirical point. The dataset construction is more careful than most: they standardize frames, filter by optical flow and scene cuts, and manually review captions and prompts.\n\nThe soft spots are substantial. Table 1 has no error bars or significance tests, so the word \"significantly\" appears a lot but is never established. The variant definitions are under-specified: they do not say exactly which modules are inserted where for each baseline. And the user-study numbers are internally inconsistent: 30 participants times 25 comparisons does not equal 1,536 responses, and the appendix says 21 pairs while the main text says 25. Minor, but it adds to the sense that the evaluation was put together in haste.\n\nThe bigger problem is the confound the stress-test note identifies. DAPE(TAV) vs TAV shows near-zero deltas, while DAPE(CCEdit) and DAPE(RAVE) show the big jumps. Since RAVE is training-free and CCEdit is trained offline, those comparisons mix the proposed modules with the mere effect of per-video fitting. Without a matched baseline that also fine-tunes on the source clip, the mechanism is unproven. They could fix this by adding a per-video full fine-tuning or LoRA baseline with the same protocol, and by reporting multi-seed means and standard deviations.\n\nBottom line: this is a believable paper with a useful benchmark, and the dual-stage observation is worth talking about. But the central empirical claim needs a serious revision. I would send it to review, partly because the dataset may be valuable to the community, but I would expect the authors to redo the comparison table. For your own work: if you are doing video editing PEFT, it is worth a skim, but I would not cite it yet.","headline":"The dual-stage PEFT recipe and the benchmark are worth a look, but the paper's central quantitative claim is confounded by per-video fine-tuning and unsupported by error bars, so the 'significantly outperforms' should be taken with skepticism.","tokens_in":21183,"tokens_out":2687,"would_cite":false,"duration_ms":26892,"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":"DAPE claims that tuning a video editor's normalization layers and then a single small visual adapter, in two separate stages, improves temporal consistency and text-video alignment for diffusion-based video editing at low training cost.","keywords":["video editing","diffusion models","parameter-efficient fine-tuning","temporal consistency","text-video alignment","norm tuning","visual adapter","video editing benchmark"],"falsifier":"Take a set of source videos, fit DAPE to one video, then apply the same fitted model to edit a different source video with the same prompt type; if the consistency and alignment gains over baselines disappear on these unseen videos, the reported improvements are memorization of the fitted clip rather than a general editing ability.","tokens_in":20005,"feed_emoji":"🎬","tokens_out":6217,"duration_ms":56490,"temperature":0.7,"pith_summary":"Video editing by diffusion models faces a trade-off: training-based editors are costly, training-free ones are inconsistent. DAPE claims to resolve this by fine-tuning very few parameters in two separate stages: first only normalization layers are adjusted to stabilize motion across frames, then a small visual adapter is trained to sharpen visual quality. The paper argues that running these two steps sequentially avoids a negative interaction observed when the modules are trained together, and shows that applying DAPE to several existing editors improves their temporal coherence and text-video alignment metrics. It also contributes a curated 232-video benchmark with manual annotations and six editing prompts per video, meant to fix imbalances in earlier evaluation sets.","feed_headline":"Two-stage fine-tuning boosts video editing consistency and alignment","feed_subtitle":"Lightweight fine-tuning beats heavier editors on coherence and prompt adherence, the authors report.","key_machinery":"Two small trainable modules carry the argument. Adjustable norm-tuning applies $\\hat{z}_t = \\gamma \\cdot \\mathit{Norm}(z_t) + \\beta + \\gamma_0 z_t$, letting the model rebalance normalized and residual features through a learnable scale $\\gamma_0$ initialized at zero. The visual adapter is a residual block of layer normalization, down-projection, a depth-wise $5\\times 5$ convolution, GELU activation, and up-projection, inserted only in the first cross-attention block of the decoder. The dual-stage training schedule first tunes norm parameters (400 timesteps), then freezes them and trains the adapter (70 timesteps), with a Huber loss replacing the usual mean-squared-error loss. DDIM inversion provides the noise initialization at inference.","core_discovery":"On the paper's own terms, DAPE establishes that a dual-stage parameter-efficient fine-tuning recipe can improve the temporal consistency and prompt alignment of text-guided video editors without full model retraining. Stage one tunes normalization affine parameters plus a learnable residual scale; stage two trains a lightweight visual adapter inserted only at the first cross-attention block of the decoder. The central empirical claim is that decoupling these two modules removes the harmful interaction that joint training produces, and that the resulting method outperforms prior methods on temporal and alignment metrics across four datasets. The paper further claims that existing video-editing benchmarks are inadequate and that its 232-video dataset, with balanced categories and standardized frame counts and resolutions, enables fairer evaluation.","pith_inferences":["If the per-video fitting is what carries the gains, DAPE is best understood as a test-time personalization tool; a held-out generalization test would clarify whether the learned norm and adapter weights transfer to new source videos.","The depth-wise 5x5 kernel in the adapter suggests local texture is what most improves perceived quality; ablating kernel size could reveal whether global structure matters at all.","Because DAPE is demonstrated on top of several different editors, the norm-tuning and adapter recipe may generalize to other temporal-attention diffusion backbones and could be tested on longer or higher-resolution videos.","The benchmark's difficulty labels could be used to identify which edit types (subject, background, event, style, combination) remain failure-prone, which the current aggregate metrics do not reveal."],"forward_implications":["Applying DAPE on top of an existing video editor improves its temporal consistency and text alignment on most metrics, with the largest gains on warp error, up to about 35 percent relative improvement.","Training the norm parameters and the adapter in one stage degrades both, supporting the paper's dual-stage schedule as the reason the components do not cancel out.","Adapter placement matters: the first decoder cross-attention block yields the best balance of smoothness and semantic alignment, while all-position or shallow placements hurt structural coherence.","The new benchmark standardizes resolution at 512x512 and frame counts at 32, 64, or 128 with balanced subject, background, and event categories, giving evaluation coverage that earlier datasets lack.","User-study rankings place DAPE first on temporal consistency, text alignment, and overall quality compared with the five baselines tested."],"supporting_citations":[{"why":"Supplies the pretrained temporal layers used as initialization and the conditioning structure reused at DAPE inference.","marker":"[12]"},{"why":"Provides a video-editing baseline and its dataset, along with the grid-concatenation and noise-shuffling conditioning scheme DAPE builds on.","marker":"[24]"},{"why":"Defines the one-shot tuning paradigm for single-video editing with temporal attention that DAPE extends.","marker":"[54]"},{"why":"Shows temporal consistency is sensitive to normalization scales in temporal layers, motivating DAPE's norm-tuning stage.","marker":"[72]"},{"why":"Supplies the DDIM inversion procedure that maps the source video into initial noise for editing at inference.","marker":"[43]"},{"why":"Provides the latent diffusion model backbone whose U-Net layers and normalization parameters DAPE tunes.","marker":"[38]"},{"why":"Introduces the adapter-tuning mechanism on which the visual adapter module is based.","marker":"[19]"},{"why":"Supplies the benchmark dataset used as one of the four evaluation sets.","marker":"[55]"},{"why":"Supplies the category scheme and video source used to construct the new DAPE dataset.","marker":"[59]"},{"why":"Establishes that normalization parameters play a role in conditional control, supporting the choice of norm layers as the tuning target.","marker":"[30]"}],"fun_headline_variants":["Dual-stage PEFT boosts video editing consistency","Two-step fine-tuning for coherent video edits","Lightweight dual adapters improve video editing","DAPE: Efficient tuning for consistent video editing","Two-stage tuning upgrades video editing alignment"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method fine-tunes on the very video clip it later edits, so the central assumption is that the fitted norm and adapter parameters learn a reusable editing behavior rather than memorizing the appearance of that clip.","fun_headline_variants_meta":{"raw":{"variants":["Dual-stage PEFT boosts video editing consistency","Two-step fine-tuning for coherent video edits","Lightweight dual adapters improve video editing","DAPE: Efficient tuning for consistent video editing","Two-stage tuning upgrades video editing alignment"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000165,"raw_usage":{"total_tokens":1219,"prompt_tokens":883,"completion_tokens":336,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":269}},"tokens_in":499,"tokens_out":336,"duration_ms":3526,"temperature":1.0,"reasoning_tokens":269,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:26:28.201476+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of source videos, fit DAPE to one video, then apply the same fitted model to edit a different source video with the same prompt type; if the consistency and alignment gains over baselines disappear on these unseen videos, the reported improvements are memorization of the fitted clip rather than a general editing ability.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides a video-editing baseline and its dataset, along with the grid-concatenation and noise-shuffling conditioning scheme DAPE builds on."}],"review_version":1}