{"id":"51ecc3d4-40df-4070-b052-8ef7c1e40d6b","arxiv_id":"2607.29627","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A single video-diffusion framework composites both static images and dynamic footage along user-defined trajectories by transporting canonical foreground latents directly into the background latent sequence.","lead":"FlexComposer inserts a static photo or an existing video clip into a background video so the inserted object follows a user-drawn path, while keeping the object's own motion. It does this by moving the object's visual features inside a diffusion model's latent space instead of adding separate control networks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Wan's VAE translation equivariance is asserted, not measured; Eq. (2)'s parameter-free latent transport may be lossy, undercutting the claimed advantage over learned adapters.","rationale":"The reader correctly identified the unverified translation-equivariance assumption as the weakest point. I agree with this assessment. The paper's headline contribution is the parameter-free latent injection, and the mechanism's validity depends on the VAE's equivariance. No part of the paper measures this. The experiments are extensive but they compare whole systems, not the fidelity of the injection; the ablation of the injection does not isolate equivariance. A direct latent-composite comparison is feasible with the described components and would settle the question. The paper's other limitations (physics, long-video inconsistency) are honestly stated and do not threaten the central claim. The evaluation issues (private V2V test set, no error bars, V2V using full background context) are concerning for the strength of the evidence, but they are secondary to the theoretical assumption; if the equivariance holds, those concerns would only affect the degree of confidence, not the mechanism. Therefore, I recommend no change to the reader's CONDITIONAL verdict, with the equivariance measurement as a necessary condition for full acceptance.","tokens_in":19565,"tokens_out":7588,"duration_ms":92653,"concrete_test":"Construct a test from the paper's own pipeline: select 20 foreground assets (static and dynamic) and 20 background videos. For each, define a set of trajectories. Build pixel-space ground-truth composites by pasting the foreground with its mask along the trajectory. Encode these with the frozen Wan VAE to get Z_gt. Then build C_fuse exactly as in Eqs. (1)-(3) using the same inputs. Report: (1) mean L2 distance between C_fuse and Z_gt normalized by ||Z_gt||, (2) the same after removing the background influence by masking to the foreground region, and (3) a diffusion rollout: feed identical noise to the trained model with condition C_fuse vs. condition Z_gt and measure output FVD and EPE. If the relative latent distance is small (e.g., <0.05) and rollout metrics are statistically indistinguishable, equivariance holds. If C_fuse differs substantially from Z_gt, the parameter-free transport","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central novelty is the parameter-free Spatial-Aware Latent Injection, whose correctness hinges on Eq. (2): copying latent features from canonical position c+δ to target u_n+δ must place the semantically correct content at each target location. This requires the Wan video VAE encoder to be translation-equivariant under integer latent shifts for arbitrary foreground content, including near-object-boundary features and over time (temporal compression f_t). Strided convolutions and patchification make exact equivariance unlikely, and Eq. (1)'s nearest-integer rounding adds additional coordinate error (up to 0.5/f_s pixels per dimension). The paper provides no measurement of this equivariance error, no analysis of how rounding affects the injected condition, and no comparison of C_fuse to the VAE encoding of a pixel-space ground-truth composite. The ablation 'w/o Latent Transport' only removes the entire transport module, so it cannot isolate whether the transport is faithful or whether the diffusion model's LoRA-finetuned priors learn to correct systematic shifts. If the latent space is not sufficiently equivariant, the 'parameter-free' mechanism is not doing what is claimed, and the SOTA quantitative results may reflect learned compensation rather than the proposed geometric transport. This is the load-bearing assumption behind the paper's core claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents FlexComposer, a video compositing framework that inserts static images or dynamic clips into background videos along user-specified 3D trajectories. The method has three components: pixel-space canonicalization (stabilization for video, expansion/noise injection for images), a `parameter-free` latent injection that copies VAE features from a canonical location to trajectory positions using assumed translation equivariance, and a synthetic-to-real curriculum. The system is built on Wan 2.1 I2V with LoRA fine-tuning. Experiments compare against I2V trajectory-control baselines, V2V compositing/editing baselines, and harmonization methods, reporting EPE, FVD, PSNR, SSIM, and VBench metrics.","tokens_in":19857,"tokens_out":7235,"duration_ms":83563,"significance":"If the central mechanism is validated, the paper offers a useful unified treatment of two previously separate settings (static and dynamic foregrounds), with an appealingly simple control mechanism. The extensive ablations and user study are strengths. However, the key novelty rests on an unverified equivariance property, and the evaluation protocol has circularity and reference-frame ambiguities. With additional analysis and more rigorous evaluation, the contribution would be publishable; in current form the headline claims are not fully supported.","major_comments":[{"comment":"The parameter-free transport operation is the paper's central novelty, but it assumes the Wan video VAE is translation-equivariant under integer latent shifts for all foreground content and across temporally compressed latent frames. Strided convolutions, patchification, and temporal compression make exact equivariance unlikely; Eq. (1) also rounds coordinates, introducing up to 0.5/f_s pixel error. The paper provides no measurement of equivariance, no comparison of C_fuse to the VAE encoding of a pixel-space ground-truth composite, and no sensitivity analysis for rounding or fast motion. Consequently the reported advantage over learned adapters may reflect LoRA-finetuned compensation rather than faithful transport. I ask for a direct equivariance test (e.g., encode shifted inputs and measure feature-shift error), a C_fuse-vs-E(pixel composite) comparison, and an analysis of how rounding","section":"§3.2, Eqs. (1)-(2)"},{"comment":"EPE is computed by tracking generated centroids with SpatialTrackerV2, and the same tracker is used to generate pseudo-ground-truth trajectories for Phase 2 and Phase 3 training data, as well as for the canonical stabilization in §3.1. This creates correlated evaluation: systematic tracker errors are shared between the training target and the evaluation measure, potentially inflating trajectory-adherence gains. Also, Tables 1-3 report point estimates only, with no confidence intervals or significance tests, and the V2V benchmark is a private 50-clip set with no release plan. Please evaluate with an independent tracker or human-annotated trajectories, release the benchmark, and report variance.","section":"§A.4, §B.2, Tables 1-3"},{"comment":"PSNR and SSIM are reported for trajectory-controlled generation, but the reference video for these metrics is not defined. In the static foreground compositing setting there is no ground-truth composite; if the reference is the unmodified background video, inserting any foreground will lower PSNR/SSIM, making the numbers uninterpretable; if a composited ground truth is used, its construction must be described. Please specify the reference and, ideally, replace these metrics with task-appropriate perceptual tests.","section":"§4.2, Tables 1-2"},{"comment":"The ablation labeled 'w/o Latent Transport' is described in the supplementary as replacing the transport module with a learned convolutional encoder, not as removing transport entirely. The main text's discussion ('sliding artifacts,' 'fail to anchor') reads as if the module were absent. As presented, the ablation cannot isolate whether the proposed transport is faithful or whether the LoRA-finetuned backbone learns corrections. Please align the label and description and add an ablation that varies only the transport operation while holding the condition signal and training protocol constant.","section":"Table 4, §C.2"}],"minor_comments":[{"comment":"'Without auxiliary learnable adapters' is potentially misleading because LoRA adapters (rank 64) are trained on the backbone. Clarify that 'parameter-free' refers only to the injection function, not the overall system.","section":"Abstract/§3.2, §A.1"},{"comment":"Define f_s, f_t and the nearest-integer symbol before using; also justify the temporal averaging over f_t frames, which may alias fast global motion.","section":"Eq. (1)"},{"comment":"The label 'w/o Noise in Static Exp.' is later called 'w/o Static Noise' in the text; use consistent terminology.","section":"Table 4"},{"comment":"Provide details of the 50-clip V2V test set (sources, trajectory distributions, masking protocol) or release it.","section":"§4.3"},{"comment":"The user study of 25 evaluators × 20 cases should report confidence intervals; 'dominant lead' overstates precision.","section":"§D"},{"comment":"The text cites Grounded-SAM2 as [35], but [35] is Segment Anything; check the citation mapping. Similarly, ensure SpatialTracker v2 citations are consistent.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"This is a technically interesting submission with a clear central idea, but the verification gap around translation equivariance and the evaluation circularity are substantial. I would only accept after these are addressed; the manuscript is not ready in current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a well-built engineering paper whose main trick—transporting canonical latent features along a trajectory by direct coordinate shifts—is elegant and probably mostly works, but the paper never tests the translation-equivariance assumption that the trick depends on. That's the thing to look at first.\n\nWhat's actually new: FlexComposer treats static and dynamic foreground assets in one framework by stabilizing assets into a canonical representation, injecting features via VAE latent transport, and training with a synthetic-to-real curriculum. That combination is not in the prior work they cite. The experiments are broad: DAVIS, MoveBench, a V2V set, harmonization, ablations, and a user study. The ablations are informative, the supplementary material is unusually complete, and the limitations section is honest about no physics and long-video instability. Credit where due.\n\nSoft spots, in order of size. First, the latent transport in Eq. (2) assumes the Wan VAE encoder is translation-equivariant under integer latent shifts. Strided convolutions and temporal compression make that only approximate, especially at object boundaries and for fast motion. The paper does not measure the equivariance error, does not analyze the rounding in Eq. (1), and does not compare the transported latent against the encoding of a pixel-space composite. So the 'parameter-free' claim is not actually evidenced. It is not fatal—the LoRA-finetuned diffusion model can learn to compensate for systematic transport errors—but it means the stated advantage over learned adapters is not established. The ablation 'w/o Latent Transport' replaces the whole transport with a learned adapter, so it cannot isolate this.\n\nSecond, the EPE metric is computed with SpatialTrackerV2, which is also used to generate pseudo-ground-truth trajectories for training data. That creates a real risk of metric bias. It would be good to see a second tracker or human-annotated trajectories. Third, standard hygiene issues: no error bars or significance tests, the V2V benchmark is a private 50-clip set, and the user study is only 25 people with modest preferences over Wan-Move. Minor in comparison.\n\nOverall: the core contribution is plausible and the paper is worth engaging. It deserves a serious referee, but the equivariance question should be a required check, not an optional one.","headline":"Clever, honest engineering paper on trajectory-guided compositing, held back by an unmeasured latent-equivariance assumption; deserves review but needs a direct equivariance check.","tokens_in":20361,"tokens_out":2322,"would_cite":false,"duration_ms":28673,"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":"FlexComposer turns video compositing into a single trajectory-guided generation step, handling both static images and pre-animated footage without explicit 3D reconstruction or auxiliary control adapters.","keywords":["video compositing","trajectory control","latent injection","translation equivariance","video diffusion","canonical foreground representation","synthetic-to-real curriculum","video harmonization"],"falsifier":"Encode a sharp object at two positions differing by one latent-grid unit; if the difference between the shifted encoding and the encoding of the shifted image is not near zero, or if decoding the transported latent blurs or smears object boundaries, the parameter-free transport fails. Additionally, running a trajectory with sub-latent-unit per-frame motion should produce a stationary or wobbling object if the nearest-integer rounding in Eq. (1) erases the motion—tracking the centroid would reveal it.","tokens_in":19426,"feed_emoji":"🎬","tokens_out":6510,"duration_ms":67258,"temperature":0.7,"pith_summary":"The paper sets out to prove that video compositing—placing an external asset into a background video—can be reframed as one trajectory-guided conditional generation task that works equally for static images and already-animated video clips. Its claim rests on three designs: a canonical foreground representation that separates the object's own motion from its global displacement, a parameter-free latent-injection mechanism that moves the object by shifting its VAE features along the user's path, and a synthetic-to-real curriculum that teaches the generative model to harmonize lighting and shadows implicitly. If these claims hold, multi-stage VFX pipelines for insertion can collapse into a single generative pass, and dynamic assets keep their intrinsic motion while following new paths—something image-animation methods cannot do. A sympathetic reader would care because this would make controlled compositing both simpler and more faithful, for product photos and for footage alike.","feed_headline":"Drop any asset onto any path in a video—no 3D stage needed","feed_subtitle":"Static images get motion; moving clips keep their own motion, with lighting and shadows adapted to the scene.","key_machinery":"The load-bearing mechanism is Spatial-Aware Latent Injection, expressed in Eq. (2): after the foreground is stabilized into a canonical latent sequence, the user's 3D trajectory is projected to 2D latent coordinates (Eq. 1) and the object's latent features are transported to each trajectory position by a direct coordinate shift, gated by a visibility mask that handles occlusions. The justification for this parameter-free transport is the translation equivariance of the video VAE latent space, which lets a shift replace any learnable adapter. Two supporting mechanisms carry the rest of the argument: the Unified Canonical Foreground Representation (motion decoupling and static expansion with G","core_discovery":"The central discovery is that a video VAE's latent space can be treated as translation-equivariant to the precision needed for direct feature transport: canonical foreground features are copied from their centered position to target trajectory coordinates via Eq. (2), with the temporal index coupled on both sides so intrinsic motion survives the move. Because the injection is a pure coordinate shift, no learnable control adapter is needed, and the paper argues this avoids the signal degradation that adapters introduce. The Unified Canonical Foreground Representation—centering dynamic videos, expanding static images with temporal noise—makes diverse inputs uniform for this same mechanism, and","pith_inferences":["Inference: If translation equivariance is as strong as the method assumes, the parameter-free injection logic should transfer to other latent video diffusion models with only backbone fine-tuning, suggesting a general recipe for trajectory-controlled editing beyond compositing.","Inference: The representation's lack of geometric warping implies a testable boundary: motion transfer quality should degrade as source and target subjects diverge in shape; a systematic shape-difference sweep would map this boundary.","Inference: The paper's stated limitation that physics are not simulated points to an immediate extension—replacing the visibility gate with simple depth ordering or a lightweight physics engine—which could remove ghosting and implausible interactions.","Inference: The reported training budget (roughly 8k steps over three curriculum stages) leaves open whether trajectory adherence scales smoothly with more real footage or denser trajectory supervision, an experiment worth running for new domains."],"forward_implications":["A single static product photo can be animated along any user-defined path while preserving identity, with the model hallucinating plausible micro-motion and view changes.","Pre-animated footage—an explosion, a fluttering wing—can be repositioned onto a new trajectory without losing its internal dynamics, a capability image-animation baselines lack.","Occlusion is explicitly controllable through the visibility gate: the asset disappears behind occluders instead of ghosting through them.","Because motion is decoupled from appearance at the representation level, the same framework supports non-pixel-aligned motion transfer from a source subject to a different target subject in a new background.","Lighting, shadows, and reflections emerge from the curriculum rather than from a dedicated harmonization stage, improving perceptual consistency over color-only harmonization methods."],"fun_headline_variants":["FlexComposer: move objects along any path, no adapters","Copy-and-paste video compositing with trajectory control","Video compositing without 3D: just a coordinate shift","FlexComposer: paste images or footage onto any trajectory","Trajectory-guided compositing: motion preserved, no 3D"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole mechanism rests on the assumption that moving a foreground's latent features by an integer offset in the latent grid places the semantically correct content at the target location, for every frame and every offset—a near-exact translation equivariance of the video VAE that the paper does not directly measure.","fun_headline_variants_meta":{"raw":{"variants":["FlexComposer: move objects along any path, no adapters","Copy-and-paste video compositing with trajectory control","Video compositing without 3D: just a coordinate shift","FlexComposer: paste images or footage onto any trajectory","Trajectory-guided compositing: motion preserved, no 3D"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000665,"raw_usage":{"total_tokens":2889,"prompt_tokens":778,"completion_tokens":2111,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":522,"completion_tokens_details":{"reasoning_tokens":2039}},"tokens_in":522,"tokens_out":2111,"duration_ms":16601,"temperature":1.0,"reasoning_tokens":2039,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T03:11:21.249069+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Encode a sharp object at two positions differing by one latent-grid unit; if the difference between the shifted encoding and the encoding of the shifted image is not near zero, or if decoding the transported latent blurs or smears object boundaries, the parameter-free transport fails. Additionally, running a trajectory with sub-latent-unit per-frame motion should produce a stationary or wobbling object if the nearest-integer rounding in Eq. (1) erases the motion—tracking the centroid would reveal it.","supporting_citations":[],"review_version":1}