{"id":"1606c5ed-d96f-4f4c-9118-7ed4213531d0","arxiv_id":"2506.22298","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"OutDreamer couples a diffusion transformer with mask-driven self-attention and a latent alignment loss to outpaint videos in a zero-shot manner, exceeding prior zero-shot baselines on standard benchmarks.","lead":"OutDreamer is a new system that creates missing video content outside the borders of an input clip using a diffusion transformer, reporting better scores than earlier zero-shot video outpainting methods on DAVIS and YouTube-VOS. The work points toward editing tools that can extend video without slow per-video fine-tuning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's SOTA claim depends on baseline numbers whose protocol is unverified; the paper admits SDM code is unreleased, yet still lists SDM in Table 1.","rationale":"The strongest claim is a comparative one: OutDreamer outperforms state-of-the-art zero-shot methods. Every quantitative support for that claim is in Table 1. The reader's concern about the SDM row is well-founded, and the manuscript itself provides the evidence: Figure 5 says SDM cannot be included because its code is not released. That makes it very unlikely the authors re-ran SDM under their own protocol; instead the numbers were taken from the M3DDM paper. A comparison that mixes protocols is not a valid comparison. The absence of variance estimates compounds the problem because the claimed advantages in PSNR are tiny (e.g., 20.30 vs 20.26 and 20.21 vs 20.20). I therefore agree with the reader's conditional posture. The architecture and ablations are reasonably described, but the central 'outperforms' claim cannot be accepted as verified until at least one baseline is rerun under the same pipeline or a protocol match is demonstrated. No change to the reader's verdict is needed: it should remain conditional, with release of evaluation code and baseline reruns as conditions.","tokens_in":14367,"tokens_out":5829,"duration_ms":62035,"concrete_test":"Ask the authors for the SDM and M3DDM checkpoints/evaluation code, or fetch the original M3DDM release, and rerun all baselines under OutDreamer's exact protocol from Section 5.1.3 and Supp. III: mask ratios 0.25 and 0.66, 256x256 resize, input blending, ShareGPT4V captions for text-conditioned methods, and FVD with 16 uniformly sampled frames. If the reproduced SDM row differs from Table 1 by more than the reported win margins (e.g., SSIM by >0.02 or FVD by >30), or if the M3DDM numbers shift similarly, the outperformance claim is an artifact of protocol mismatch. If the original papers do not match on mask ratio or FVD sampling, that mismatch is sufficient to conclude the comparison is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on Table 1, so the load-bearing condition is that every baseline entry was produced under the same evaluation protocol as OutDreamer. Section 5.1.3 specifies OutDreamer's pipeline: average of 0.25 and 0.66 mask ratios, resize to 256x256, input blending, ShareGPT4V captions, and 16-frame FVD. The paper never states that Dehan, SDM, or M3DDM were re-run under that pipeline. This matters because Figure 5's caption explicitly says 'SDM is not included since its code is not released,' yet Table 1 lists SDM scores for both datasets; the likely source is the M3DDM paper, whose mask ratios, resolution, blending, and FVD sampling need not match. The same concern applies to M3DDM if its numbers were copied rather than recomputed. In addition, no error bars or significance tests accompany the numbers, so the PSNR gaps of 0.01-0.04 dB are indistinguishable from run-to-run noise. The SOTA claim is therefore not yet verifiable; this is a fairness-of-comparison problem, not a flaw in the architecture itself.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes OutDreamer, a zero-shot video outpainting framework built on a diffusion transformer backbone (Open-Sora-Plan). The method has two main branches: an efficient video control branch that injects masked video latents into early DiT blocks, and a conditional outpainting branch with mask-driven self-attention that scales attention keys according to the mask. A latent alignment loss aligns per-frame latent means and variances, and long videos are generated iteratively with a cross-video-clip refiner that performs mean-variance alignment and histogram matching. The authors report quantitative comparisons on DAVIS and YouTube-VOS against Dehan, SDM, M3DDM, and MOTIA, and present ablations for the latent alignment loss, mask-driven self-attention, the refiner, and the mask downsampling function. The central claim is that OutDreamer outperforms existing zero-shot video outpainting methods on standard benchmarks.","tokens_in":14684,"tokens_out":5527,"duration_ms":62057,"significance":"If the reported numbers are reliable, OutDreamer is a useful step toward DiT-based video outpainting: the control-branch injection at early DiT blocks, the mask-driven key scaling, and the cross-video-clip refiner are reasonable and the ablations are internally consistent. The paper also provides explicit loss equations and pseudocode for the refiner, which aids reproducibility. However, the headline claim rests on a benchmark comparison whose baseline protocol is not established, and the PSNR margins over the strongest baseline are tiny (0.04 dB and 0.01 dB). The architecture is promising, but the empirical support for state-of-the-art status needs to be verified with a fairer protocol and statistical grounding.","major_comments":[{"comment":"The evaluation protocol for OutDreamer is described in detail (average of 0.25 and 0.66 mask ratios, resizing to 256×256, input blending, ShareGPT4V captions, and 16-frame FVD sampling), but the paper never states that Dehan, SDM, or M3DDM were rerun under the same protocol. The contradiction with Figure 5 is particularly important: its caption says \"SDM is not included since its code is not released,\" yet Table 1 lists SDM scores for both datasets. If the SDM and M3DDM numbers are copied from the M3DDM paper under different mask ratios, resolutions, blending, or FVD sampling, the comparison is not fair and the claimed state-of-the-art status is not verifiable. The authors should rerun all baselines under the exact protocol used for OutDreamer, or if that is impossible, state the exact source of each baseline number and justify why the protocols are equivalent.","section":"§5.1.3, Table 1, Figure 5"},{"comment":"No error bars, confidence intervals, or significance tests accompany the quantitative results. The PSNR differences over M3DDM are 0.04 dB on DAVIS and 0.01 dB on YouTube-VOS, which are within the run-to-run variability expected from stochastic diffusion sampling. Additionally, the reported scores are averaged over mask ratios 0.25 and 0.66; without per-ratio breakdowns or multiple-seed evaluation, the statement that OutDreamer \"outperforms them in all the scores\" is not supported, particularly for PSNR.","section":"Table 1, §5.1.2"},{"comment":"The training setup is not described with sufficient detail to assess the method or reproduce it. The paper specifies loss functions, hyperparameters β and T_latent, and a mask-ratio sampling rule, but it never states which video dataset or datasets were used for training, how many videos, at what resolution and frame length, or for how many training iterations. Since OutDreamer is a learned zero-shot method, the training data is a core part of the method and must be reported.","section":"§4.4, §5.1.3, Supplementary III"},{"comment":"The long-video evaluation is underspecified. The paper states that 20 long videos from Pexels were used, videos were generated with varying frame lengths at mask ratio 0.5, and scores were computed for the last clip, but it does not describe how ground truth is obtained, whether the outpainting target regions are synthetically masked versions of the original videos, how the 20 videos are split, or how the scores are aggregated. The refiner ablation in Table 4 therefore lacks enough protocol detail to be independently reproduced or interpreted.","section":"§5.4, Table 4"}],"minor_comments":[{"comment":"The hyperparameter γ in the mask-driven self-attention is never given a value; please report it together with β and T_latent.","section":"Eq. (3), Supplementary III"},{"comment":"The first row of the ablation table has no method label; it should be labeled \"Baseline\" for clarity.","section":"Table 3"},{"comment":"The advanced version of the efficient video control branch is described only through Figure 2 and a short sentence; an explicit formulation analogous to Eq. (2) would clarify how the last K frames from the previous clip are encoded and injected.","section":"§4.5.1, Figure 2"},{"comment":"The figure reports \"maximum normalization with proportional scaling,\" which makes the y-axis values nonstandard and hard to interpret; please show raw metric values or describe the normalization exactly, and consider adding error bars.","section":"Figure 7"},{"comment":"The notation g_t is written as \"gt\" in the equation; please use a subscript for readability and define the range of t relative to the total number of diffusion timesteps.","section":"Eq. (6)"}],"recommendation":"major_revision","confidential_remarks":"The main obstacle to acceptance is the verifiability of the baseline comparison in Table 1. If the authors can rerun the baselines under the exact evaluation protocol, or convincingly justify that copied numbers are protocol-equivalent, and add error bars or a significance analysis, the paper could be suitable for publication. The architecture and ablations are otherwise reasonable, and there is no sign of circularity in the evaluation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read it. The architecture work is legit. OutDreamer is a DiT-based zero-shot outpainting method; I don't know of an earlier one in the cited literature. The three new pieces—the ControlNeXt-style early control branch, mask-driven self-attention (Eq. 3), and the latent alignment loss (Eq. 5)—are sensibly motivated and individually ablated. The supplementary gives pseudocode for the refiner, hyperparameters, and even an ablation of the mask downsampling function. That is more than most vision papers ship.\n\nThe soft spot is exactly where the stress test points: Table 1. The paper lists SDM scores on both DAVIS and YouTube-VOS, but Figure 5's caption says SDM is not included because its code isn't released. So where do those numbers come from? Almost certainly from the M3DDM paper, and the authors never state that Dehan, SDM, or M3DDM were re-run under OutDreamer's protocol (average of 0.25 and 0.66 masks, 256x256 resize, blending, ShareGPT4V captions, 16-frame FVD). If the baselines are copied, the comparison isn't fair and the claimed SOTA on PSNR is a wash anyway: 0.04 dB and 0.01 dB over M3DDM. With no error bars or significance tests, those differences are noise. The same protocol doubt applies to M3DDM itself if it wasn't rerun.\n\nThat said, the internal evidence is consistent. The ablations (LAL+MSA vs baseline, refiner vs no refiner) show monotone improvements on all four metrics, and the qualitative examples look better than Dehan. So I don't think the architecture is broken; the evaluation is just not yet verifiable. The training data is also not described, which matters for reproducibility, and there's no code release. Minor: the long-video benchmark is a self-built Pexels set with no baseline comparison.\n\nWho is this for? Anyone working on DiT-based controllable video generation, or on outpainting. It's a genuinely useful method paper. It deserves a serious referee, but the referee should insist on baseline protocol verification or at least the authors disclosing exactly where each Table 1 number came from, plus error bars over multiple seeds. I'd recommend major revision before acceptance, not desk rejection.","headline":"Solid DiT-based zero-shot video outpainting architecture with real ablations; the SOTA claim rests on unverified baseline numbers, so the evaluation needs a serious pass before this is citable as a benchmark win.","tokens_in":15199,"tokens_out":3436,"would_cite":true,"duration_ms":33219,"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":"OutDreamer, a zero-shot diffusion-transformer video outpainting method, reports better SSIM, LPIPS, and FVD than existing zero-shot baselines on DAVIS and YouTube-VOS, with PSNR equal or better, and parity with a one-shot method that…","keywords":["video outpainting","diffusion transformer","zero-shot video generation","mask-driven self-attention","latent alignment loss","long video generation","temporal consistency"],"falsifier":"Run the released masked 3D diffusion code (and the non-masked variant if a checkpoint can be obtained) through OutDreamer's exact evaluation pipeline—same mask ratios 0.25 and 0.66, 256x256 resizing, 29-frame clips, 16-frame FVD sampling, and same captions—and check whether the resulting scores reproduce Table 1; if the baseline numbers change materially, the claimed lead may not hold.","tokens_in":14183,"feed_emoji":"🎬","tokens_out":5796,"duration_ms":57546,"temperature":0.7,"pith_summary":"The paper introduces OutDreamer, a video outpainting method that extends video frames beyond their borders without any per-video fine-tuning. It claims that a diffusion transformer backbone, fed with masked-video conditions through a lightweight control branch and a mask-driven self-attention layer, generates visually realistic and temporally consistent outpainted content. The authors report that this zero-shot approach beats current zero-shot baselines on DAVIS and YouTube-VOS on SSIM, LPIPS, and FVD, and reaches parity with a one-shot method that is allowed to adapt to each video. If true, this would make high-quality video outpainting practical for both short and long clips without costly per-video adaptation.","feed_headline":"OutDreamer tops zero-shot video outpainting on two benchmarks","feed_subtitle":"A diffusion-transformer method leads on SSIM, LPIPS, and FVD without per-video fine-tuning.","key_machinery":"The load-bearing mechanism is the mask-driven self-attention of Eq. (3): $$ \\mathrm{Attn}(Q,K,V) = \\mathrm{softmax}\\left(\\frac{Q(K \\cdot (1+\\gamma F_s(m)))^T}{\\sqrt{d_k}}\\right) V, $$ where $F_s(m)$ maps the mask to per-key scaling factors in $[-1,1]$, so patches in the known region attract more attention during denoising. Around this sit two supporting pieces: an efficient control branch that extracts the masked-video latent and adds its features into the first DiT block (Eq. (2)), and a latent alignment loss that penalizes per-frame mean and variance differences between predicted and ground-truth latents (Eq. (5)). A cross-video-clip refiner, doing mean-variance alignment and histogram matching stitch, keeps iteratively generated long clips consistent.","core_discovery":"On its own terms, the paper establishes that replacing a U-Net denoiser with a diffusion transformer, injecting masked-video conditions right after the first transformer block, and steering self-attention with mask information yields zero-shot video outpainting that is more realistic and temporally consistent than existing zero-shot approaches. In the reported tables, OutDreamer leads the zero-shot baselines on DAVIS and YouTube-VOS for SSIM, LPIPS, and FVD, and its PSNR is equal to or better than theirs; against a one-shot method that fine-tunes per video, it achieves comparable scores without any adaptation. The ablation study attributes these gains to the latent alignment loss and the mask-driven self-attention layer, and to the cross-video-clip refiner for long sequences.","pith_inferences":["If the protocol caveat about the non-masked 3D diffusion baseline is resolved, the architecture is likely transferable to other masked video generation tasks such as object removal or region-based video editing, because the mask-driven attention and early condition injection are not outpainting-specific.","The cross-video-clip refiner relies on hand-crafted color statistics; a learned refiner trained to minimize inter-clip discontinuities could plausibly push long-video consistency further than the reported histogram matching.","The paper evaluates temporal consistency only through FVD and qualitative inspection; a direct temporal-coherence metric, such as inter-frame warping error or flicker measurement, could test whether the claimed frame-to-frame consistency is as strong as stated.","Because training uses mask ratios between 0.1 and 0.8, performance at more extreme masks (for example 0.9) is an open question that would tell whether the method scales to near-total outpainting."],"forward_implications":["If the reported numbers hold, zero-shot video outpainting does not need a U-Net backbone: the diffusion transformer version beats the U-Net-based zero-shot baselines on SSIM, LPIPS, and FVD in the paper's tables.","The mask-driven self-attention and latent alignment loss each add measurable quality; Table 3 shows the full model scores best, so future outpainting systems can adopt these two modules independently.","Long-video outpainting can be done iteratively, clip by clip, with overlapping condition frames and a refiner; Table 4 shows the refiner improves all metrics on 315-frame outputs.","Zero-shot outpainting can reach roughly the same quality as one-shot adaptation that fine-tunes per video, without the per-video cost."],"supporting_citations":[{"why":"supplies the five-stage flow and background estimation baseline that OutDreamer must beat.","marker":"[7]"},{"why":"supplies both the masked and non-masked 3D diffusion baselines and the numbers used in Table 1.","marker":"[9]"},{"why":"the diffusion transformer backbone that replaces the U-Net.","marker":"[26]"},{"why":"the early condition-injection strategy after the first transformer block that OutDreamer adapts.","marker":"[27]"},{"why":"the pre-trained text-to-video model used as the base for fine-tuning.","marker":"[15]"},{"why":"the one-shot method that fine-tunes per video and serves as the adaptation baseline.","marker":"[36]"}],"fun_headline_variants":["Diffusion transformer outpaints video zero-shot, beats SOTA","OutDreamer: zero-shot video outpainting with DiT leads benchmarks","Mask-driven attention boosts zero-shot video outpainting","OutDreamer sets new zero-shot video outpainting records","Zero-shot video outpainting: DiT-based method wins on metrics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The state-of-the-art claim assumes the SDM scores in Table 1 came from the same evaluation protocol as OutDreamer's, yet the paper says SDM's code is not released and excludes it from qualitative comparisons.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion transformer outpaints video zero-shot, beats SOTA","OutDreamer: zero-shot video outpainting with DiT leads benchmarks","Mask-driven attention boosts zero-shot video outpainting","OutDreamer sets new zero-shot video outpainting records","Zero-shot video outpainting: DiT-based method wins on metrics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000721,"raw_usage":{"total_tokens":3221,"prompt_tokens":917,"completion_tokens":2304,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":2215}},"tokens_in":533,"tokens_out":2304,"duration_ms":16140,"temperature":1.0,"reasoning_tokens":2215,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:06:55.590229+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the released masked 3D diffusion code (and the non-masked variant if a checkpoint can be obtained) through OutDreamer's exact evaluation pipeline—same mask ratios 0.25 and 0.66, 256x256 resizing, 29-frame clips, 16-frame FVD sampling, and same captions—and check whether the resulting scores reproduce Table 1; if the baseline numbers change materially, the claimed lead may not hold.","supporting_citations":[{"cited_title":"Complete and temporally consistent video out- painting","cited_arxiv_id":null,"evidence_quote":"supplies the five-stage flow and background estimation baseline that OutDreamer must beat."},{"cited_title":"Hierar- chical masked 3d diffusion model for video outpainting","cited_arxiv_id":null,"evidence_quote":"supplies both the masked and non-masked 3D diffusion baselines and the numbers used in Table 1."},{"cited_title":"Open-sora-plan, 2024","cited_arxiv_id":null,"evidence_quote":"the pre-trained text-to-video model used as the base for fine-tuning."},{"cited_title":"Be-your-outpainter: Mastering video outpainting through input-specific adaptation","cited_arxiv_id":null,"evidence_quote":"the one-shot method that fine-tunes per video and serves as the adaptation baseline."}],"review_version":1}