{"id":"e730ba51-7d65-42c8-a599-5075a7ac0b21","arxiv_id":"2511.17185","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"PostCam generates new videos from a reference video along user-specified camera trajectories using a query-shared cross-attention that fuses pose data and rendered frames, improving control precision and detail preservation.","lead":"PostCam is a system that takes a source video and a target camera path, then generates a new video that follows that path while keeping the original scene's details. Its key idea is a shared attention mechanism that mixes the camera position data with a rough 3D render of the scene so the model can be both accurate and visually faithful.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"VGGT-based camera-accuracy metric may be biased toward appearance preservation, so the >20% control-precision claim is not yet established.","rationale":"The reader's weakest_assumption identifies the VGGT-based evaluation as the most load-bearing concern, and my analysis confirms it. The paper's internal ablations (Tables 3–5) provide coherent evidence for the mechanism relative to its own baselines, and the qualitative results are plausible. However, the headline claim of 'over 20% improvement in camera-control precision' is entirely dependent on the VGGT metric. Since VGGT is a learned estimator whose outputs can depend on appearance, and PostCam is specifically designed to preserve appearance, a systematic measurement bias is a real risk. The paper gives no validation of VGGT on the test distribution and no alternative estimator. This is not an internal inconsistency but an external-validity gap. The reader's CONDITIONAL verdict is appropriate: the claim is plausible but not fully supported until the metric is validated. Hence I do not move the verdict. The concrete test I propose would settle whether the measured improvements are genuine or an artifact of measurement.","tokens_in":12390,"tokens_out":6963,"duration_ms":66548,"concrete_test":"Recompute RotErr/TransErr for all methods in Tables 1 and 2 using an independent trajectory estimator (e.g., MASt3R or DUSt3R, or a dynamic-scene SLAM variant) and compare the relative ranking of PostCam vs. ReCamMaster. Additionally, apply VGGT to the synthetic source videos (whose camera trajectories are known exactly) and report its RotErr/TransErr; then check whether VGGT's error is correlated with video appearance quality (e.g., FID against source). If PostCam's relative improvement over ReCamMaster persists with the alternative estimator, and VGGT error on known-pose videos is low and uncorrelated with FID, the concern is mitigated; if not, the central camera-control claim needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim—over 20% improvement in camera-control precision—rests on RotErr/TransErr computed by estimating camera trajectories of generated videos with VGGT, a learned feed-forward reconstruction model. On the real-world OpenVid set, there is no ground-truth trajectory; the metric compares VGGT estimates to the input trajectory. On the synthetic set, the ground-truth trajectory of the generated video is also unknown, so VGGT is again the proxy. VGGT's pose estimates depend on image content: videos that retain more source-appearance details (e.g., texture, sharp edges, recognizable objects) are likely reconstructed more accurately. PostCam is explicitly designed to preserve such details, and the paper's own FID/FVD results confirm it achieves better appearance fidelity than baselines. Thus, lower measured trajectory error for PostCam could partially reflect VGGT's preference for appearance-preserving videos rather than superior camera control. The paper does not report any validation of VGGT on the test distribution (e.g., applying VGGT to synthetic source videos with known poses and measuring its error), nor does it compare with an independent trajectory estimator. Without such a check, the headline improvement in camera-control precision is not robust to measurement bias.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PostCam, a framework for camera-controllable novel-view video generation from a source video and a target 6-DoF trajectory. Its central contribution is a query-shared cross-attention module that jointly injects camera-pose embeddings and a rendered-video condition into a frozen Wan2.1 diffusion transformer, alongside a two-stage training strategy (pose-only first, then joint with rendered video). The method is evaluated on real-world OpenVid clips and a synthetic dataset against ReCamMaster, TrajectoryCrafter, and DaS, reporting improvements in RotErr/TransErr, VBench video-quality scores, and FID/FVD. The authors claim over 20% improvement in camera-control precision and view consistency.","tokens_in":12616,"tokens_out":6119,"duration_ms":55708,"significance":"If the empirical results are reliable, PostCam is a useful empirical contribution: it shows that a simple shared-query attention mechanism over two heterogeneous control modalities (numerical poses and rendered visuals) can improve trajectory control without added input complexity or heavy parameter overhead. The paper includes systematic ablations (Tables 3–5) isolating the conditioning strategy, attention-injection mechanism, and training order, which is a strength and aids reproducibility. However, the central >20% claim rests on a camera-accuracy metric that uses VGGT pose estimates without validation on known trajectories, and all quantitative comparisons are single point estimates with no uncertainty quantification. These issues must be addressed before the claim is convincing.","major_comments":[{"comment":"The headline camera-control claim depends entirely on VGGT, a learned feed-forward reconstruction model, to estimate trajectories of generated videos. On the real-world OpenVid set there is no ground-truth trajectory; on the synthetic set ground-truth trajectories are known, but the paper does not report VGGT's pose-estimation error on those GT videos. Since VGGT estimates can be biased toward videos that preserve sharp appearance details—which PostCam is designed to do—the reported RotErr/TransErr improvements could partly reflect measurement bias rather than true control precision. The manuscript mentions a supplementary comparison of trajectory estimators but does not include it in the main text. Please provide a validation table showing VGGT's RotErr/TransErr on the synthetic GT videos (or on a held-out set with known trajectories), and/or an independent trajectory estimator, to rule","section":"§4.1, Evaluation Metrics (Camera Accuracy)"},{"comment":"All metrics are single point estimates with no standard deviations, confidence intervals, or significance tests. The test sets are small (100–1000 samples), and the claimed improvements—e.g., Table 1 RotErr 0.0649→0.0501, FID 85.68→67.20; Table 2 FID 103.74→80.29—could fall within run-to-run variance. The paper's conclusion uses the word \"significant\" without statistical support. Please report mean±std over multiple training runs or seeds, or at least provide bootstrap confidence intervals for the key comparisons, so the reader can assess whether the >20% margins are robust.","section":"Tables 1–5 (all quantitative comparisons)"},{"comment":"The paper defines \"view consistency\" as FID and FVD computed between the generated video and the source (real-world set) or GT target (synthetic set). FID/FVD are distribution-level fidelity measures, not view-consistency metrics; they do not measure the spatial/temporal correspondence of novel views. The repeated claim of \"over 20% improvement in view consistency\" is therefore based on a conflation of two distinct notions. Either rename the metric to \"distribution fidelity\" or use an established view-consistency metric (e.g., feature-based cross-view matching, estimated depth/point-cloud consistency) to support the claim.","section":"§4.1, View Consistency metric"}],"minor_comments":[{"comment":"\"hardly to preserve fine visual details\" should be reworded to \"often fails to preserve\" or \"hardly preserves.\"","section":"§1"},{"comment":"The method name is inconsistently spelled: \"ReCamMaster\" appears in most places but \"RecamMaster\" appears in Section 4.1 and elsewhere. Please unify.","section":"Throughout"},{"comment":"The equations contain formatting artifacts (e.g., \"token-dimp\", \"softmax\" line breaks). Please clean up the LaTeX/PDF rendering.","section":"§3.4, Eqs. (4)–(6)"},{"comment":"The caption contains a stray line of \"dddddddddddddddddd\" that appears to be a formatting artifact and should be removed.","section":"Figure 5 caption"},{"comment":"The phrase \"over 20% significant improvement\" is ambiguous: \"significant\" may be interpreted statistically. Since no significance tests are reported, consider rewording to \"over 20% improvement\" or provide statistical evidence.","section":"Abstract and Conclusion"},{"comment":"DaS is an image-to-video model, not a video-to-video recapture model. Including it in the main comparison is somewhat unfair because it is not designed to preserve source-video content. If kept, the paper should clearly mark which baselines are I2V and which are V2V, or move DaS to a supplementary comparison.","section":"§4.2, Baselines"}],"recommendation":"major_revision","confidential_remarks":"The paper is well-structured and the architectural idea is plausible, but the empirical validation has a load-bearing measurement-bias concern: the camera-accuracy metric relies on VGGT without demonstrating its reliability on known trajectories, and all results lack uncertainty quantification. The VGGT validation and error bars are essential. If the authors can show that VGGT is unbiased with respect to appearance preservation and that the improvements are statistically meaningful, this could become a solid contribution. The FID/FVD \"view consistency\" conflation also needs correction."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the paper. It's a real architectural contribution, not a curve-fit. The query-shared cross-attention, where one query attends to concatenated key/value pairs from pose embedding and rendered video, is a cleaner way to fuse heterogeneous control signals than the additive or channel-concat tricks in MVGenMaster and Uni3C. The two-stage training—pose first, then joint—is also well motivated and the ablations back it up: separate cross-attention layers or independent K/V injection do measurably worse, as do one-stage or render-first schedules. These ablations compare apples to apples, and the gaps are large enough that I'd be surprised if they were pure noise. The writing is honest about convergence problems with naive fusion, and the qualitative frames show a real robustness advantage in dynamic scenes.\n\nNow the soft spots. The headline \"over 20%\" camera-accuracy improvement rests on a metric I don't trust yet. The paper estimates trajectories of generated videos with VGGT, a learned feed-forward model, because COLMAP fails on dynamic scenes. But there's no ground-truth trajectory for any generated video, real or synthetic, so RotErr and TransErr are always measured relative to VGGT's estimates. VGGT's pose output depends on image content: videos that keep sharp edges and recognizable objects are easier to reconstruct, and PostCam is explicitly designed to preserve those details. The paper's own FID/FVD numbers confirm it does. So part of PostCam's lower trajectory error could be VGGT rewarding appearance preservation, not superior camera control. The paper doesn't validate VGGT on known-pose videos or compare against an independent estimator, and the supplementary comparison of trajectory methods is not included here.\n\nMinor but related: no error bars or significance tests anywhere. Single point estimates across all tables. Given that real-world VGGT estimates are noisy, individual RotErr numbers like 0.0501 vs 0.0649 could shift under re-runs, even if the trend holds. Code and test sets are promised but not yet available, which makes it hard to verify the exact numbers.\n\nThese are fixable issues, not a fatal flaw. The mechanism is plausible, the ablations are honest, and the qualitative results are compelling. I'd send this to peer review, with the condition that the authors validate their pose-estimation pipeline and release code/data. For my own work I'd cite the architecture but avoid relying on the specific numerical claims. Good paper to bring to reading group if you want to discuss evaluation bias in video generation metrics.","headline":"PostCam's shared-query cross-attention and two-stage training look like a genuine step forward for camera-controllable video generation, but the VGGT-based camera metric means the \"over 20%\" improvement claim isn't established yet.","tokens_in":13153,"tokens_out":2915,"would_cite":true,"duration_ms":30280,"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":"Fusing pose and rendered video through one shared query sharpens camera control in novel-view video generation by over 20%.","keywords":["novel-view video generation","camera-controllable video","video diffusion","cross-attention","6-DoF pose conditioning","rendered video conditioning","dynamic scenes","two-stage training"],"falsifier":"On synthetic videos with known camera paths, compute RotErr and TransErr both from the ground-truth poses and from the feed-forward estimator, and check whether the estimator's errors correlate with how much source-appearance detail a method preserves; if the >20% advantage shrinks under ground-truth pose comparison, or if the estimator's error is larger for outputs that discard source detail, the central claim fails.","tokens_in":12266,"feed_emoji":"🎥","tokens_out":5540,"duration_ms":44341,"temperature":0.7,"pith_summary":"The paper argues that precise camera control in novel-view video generation comes from deeply aligning two control signals — numeric 6-DoF camera poses and rendered proxy frames — rather than from adding more inputs. It introduces PostCam, which injects both modalities through a single query-shared cross-attention module, so the denoising process can pick the most relevant cues from each. Two-stage training first teaches the model camera motion from poses alone, then refines it with rendered visuals, avoiding the convergence problems that arise when both signals enter at once. The claimed payoff: over 20 percent better trajectory precision and view consistency than existing methods on real and synthetic dynamic scenes, with fine visual details like faces and hands preserved. If true, the work gives videographers a practical way to re-shoot camera moves after capture without degrading the picture.","feed_headline":"One shared query sharpens camera control in video generation by 20%","feed_subtitle":"Fusing camera poses with rendered frames in a single cross-attention keeps fine details while following user trajectories.","key_machinery":"Query-Shared Cross-Attention: a cross-attention layer where the noised-latent query is shared, while keys and values from the 6-DoF pose embedding and the rendered video are concatenated along the token dimension before the softmax. Because the same query must attend to both modalities, the model is forced to distill common camera-motion cues and suppress modality-specific noise; the zero-initialized projector and the two-stage pose-first training schedule (pose-only first, pose-plus-render second) are what make this alignment learnable.","core_discovery":"PostCam claims that the long-standing trade-off in novel-view video generation — pose control without visual detail versus rendering guidance that breaks under depth errors — can be dissolved by modality alignment rather than more input complexity. In each transformer block, a single query attends jointly to keys and values from the camera-pose embedding and from a downsampled rendered video, letting the model weigh which signal matters for each token during denoising. The rendered stream's job is narrowed to refining motion accuracy and visual fidelity, which is why the model still produces correct motion when depth maps are so bad that the rendered proxy collapses. Combined with a pose-fir","pith_inferences":["The paper measures real-scene trajectory accuracy with a learned feed-forward estimator because classic structure-from-motion fails in dynamic scenes; if that estimator rewards outputs that retain source appearance, the reported >20% precision gain could be partly a measurement artifact rather than true camera control. A synthetic ground-truth trajectory test would settle this.","The shared-query design is a general recipe for fusing heterogeneous conditioning signals, so it likely transfers to other multi-modal controls such as object motion + camera, or depth + text, in video diffusion.","Downsampling the rendered video works because the shared query learns to read coarse visual cues; an appealing extension is to make the render resolution an adjustable knob trading control granularity against compute.","If estimator bias is real, the field's standard evaluation of dynamic-camera video generation needs a benchmark with ground-truth trajectories, e.g. ray-traced synthetic scenes, rather than proxy estimation."],"forward_implications":["Camera trajectories of captured videos can be edited after the fact along arbitrary user paths while preserving high-fidelity appearance, including in dynamic scenes.","Because the rendered video is only a coarse motion-and-appearance prior, the method tolerates poor monocular depth; severe distortion in the render does not break the generated motion.","The added trainable parameters are small — a camera encoder, cross-attention, and projector — so the approach runs on a lightweight 1.3B-parameter backbone and is cheap to deploy.","The two-stage recipe gives a concrete answer to why naive multi-modal conditioning fails: simultaneous injection of pose and render signals delays convergence, while pose-first learning avoids the conflict.","Both camera-parameter (rotation-translation) and Plücker-coordinate pose representations work through the same shared-query design, so the mechanism is not tied to a specific pose encoding."],"fun_headline_variants":["Shared query fuses pose and render for 20% better camera precision","Single query aligns pose and rendered frames, cutting camera errors by 20%","PostCam's query-shared attention unites pose and render for precise video","One query to merge pose and render: 20% better video trajectories","PostCam: one shared query sharpens camera-controlled video generation"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The strong empirical claim depends on the assumption that the learned feed-forward trajectory estimator used on real videos is an unbiased measure of true camera motion, even though no ground-truth trajectory exists there.","fun_headline_variants_meta":{"raw":{"variants":["Shared query fuses pose and render for 20% better camera precision","Single query aligns pose and rendered frames, cutting camera errors by 20%","PostCam's query-shared attention unites pose and render for precise video","One query to merge pose and render: 20% better video trajectories","PostCam: one shared query sharpens camera-controlled video generation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001181,"raw_usage":{"total_tokens":4695,"prompt_tokens":701,"completion_tokens":3994,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":445,"completion_tokens_details":{"reasoning_tokens":3912}},"tokens_in":445,"tokens_out":3994,"duration_ms":30921,"temperature":1.0,"reasoning_tokens":3912,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T20:56:51.006568+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On synthetic videos with known camera paths, compute RotErr and TransErr both from the ground-truth poses and from the feed-forward estimator, and check whether the estimator's errors correlate with how much source-appearance detail a method preserves; if the >20% advantage shrinks under ground-truth pose comparison, or if the estimator's error is larger for outputs that discard source detail, the central claim fails.","supporting_citations":[],"review_version":1}