{"id":"41e0880b-1877-4e6c-b005-ed318e38f60f","arxiv_id":"2505.22944","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"ATI injects user-drawn point trajectories as soft Gaussian feature masks into a pretrained image-to-video diffusion model, enabling unified camera, object, and local motion control.","lead":"This paper introduces ATI, a system that lets users control AI-generated videos by drawing point trajectories on a single starting image. It claims to unify camera movement, object motion, and local deformations in one lightweight module that works across different video generation backbones.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation circularity: the same TAP-Net used to generate training labels is used as the evaluation tracker, and no baselines are compared, so the claimed 'superior performance' is not quantitatively supported.","rationale":"The reader correctly identified the absence of baseline comparisons and the self-referential metric as problems in the rationale, but their stated weakest assumption was the sufficiency of 2D trajectories for camera control. I agree that the camera limitation is real—the paper itself notes in Figure 5 that planar trajectories yield static content, and this tempers the 'dynamic viewpoint changes' claim. However, I find the evaluation circularity to be the more load-bearing concern because it affects every quantitative claim in the paper, including the core assertion of superiority over prior art. Even if the camera issue were resolved (e.g., by adding depth-aware trajectory encoding), the paper still would not have shown that ATI outperforms other methods without a fair, non-circular comparison. The reader's verdict of CONDITIONAL is appropriate: the method is plausible and qualitatively compelling, but the quantitative evidence must be redone with an independent tracker and at least one baseline before the central claim can be accepted. My concern does not shift the verdict; it sharpens the conditions under which the paper should be accepted. I chose 'partial' agreement because the reader's weakest_assumption field points to the trajectory/depth limitation, while my primary concern is the evaluation methodology, although the reader's rationale also flags the same evaluation issue.","tokens_in":8535,"tokens_out":4018,"duration_ms":44909,"concrete_test":"Run the same 100-pair test set used in Section 4.3 through an independent point tracker (e.g., CoTracker2 or PIPs) and recompute Acc@0.05, Acc@0.01, and appearance rate. Additionally, evaluate at least one prior trajectory/flow-based method (e.g., DragAnything or Motion-I2V) and one commercial product under the identical protocol, ideally with human raters to confirm whether specified trajectories are followed. If the metrics degrade substantially with the independent tracker, or if a baseline matches or exceeds ATI, then the claimed superiority is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of 'significantly better controllability and visual quality compared to prior approaches and commercial solutions' rests entirely on the quantitative results in Table 1, but that table compares only ATI with Seaweed-7B against ATI with Wan2.1-14B; no prior method or commercial product is evaluated. The metrics themselves are circular: Section 3.4 generates training trajectory labels with TAP-Net, and Section 4.3 evaluates ATI by tracking user-specified points in generated videos with the same TAP-Net and measuring agreement with the user trajectories. The model is thus trained to satisfy TAP-Net's tracking behavior on real videos, so high Acc@0.05/0.01 may reflect TAP-Net's inductive biases (e.g., where it tends to track, how it handles occlusion) rather than genuine point-level correspondence in the generated video. This concern is not merely stylistic; it directly undermines the only quantitative evidence for the paper's headline superiority claim. The paper even states in Section 4.2 that purely planar trajectories may yield static content, but this self-reported limitation is secondary to the fact that no non-circular, comparative evaluation exists. For the central claim to hold, the metrics must be robust to the choice of tracker and the comparisons must include at least one alternative approach.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ATI, a trajectory-conditioned image-to-video generation framework. It samples latent feature vectors at user-selected keypoints in the first frame and injects them into later frames through Gaussian spatial masks, with a tail-dropout regularizer to avoid spurious occlusions when trajectories terminate early. Training data are generated by tracking uniformly sampled points on 2.4M clips with TAP-Net, and the module is fine-tuned on Seaweed-7B and Wan2.1-14B. The paper claims unified control of camera, object, and local motion through point trajectories, with superior controllability and visual quality relative to prior methods and commercial tools.","tokens_in":8788,"tokens_out":4633,"duration_ms":42246,"significance":"If the claims were supported, the contribution would be valuable: a single lightweight module that controls camera, object, and local motion through point trajectories in pretrained I2V models would simplify workflows and improve composability. The Gaussian feature injection and tail-dropout idea are simple and plausible, the training-data scale is substantial, and demonstrations on two backbones suggest compatibility. However, the current experimental evidence does not establish the headline superiority claims: the only quantitative table compares two ATI variants, and the evaluation metric shares its tracker with the label generator. The conceptual and methodological contribution is nonetheless a reasonable starting point for a stronger evaluation.","major_comments":[{"comment":"The quantitative evaluation in Table 1 compares only ATI with Seaweed-7B against ATI with Wan2.1-14B, yet the abstract and introduction claim 'superior performance' and 'significantly better controllability and visual quality compared to prior approaches and commercial solutions.' With no baseline method or commercial system measured, this central claim is unsupported by the reported numbers; the authors should add comparisons against at least one trajectory-based prior method (e.g., Motion Prompting, ReVideo, Trailblazer) and one commercial tool, together with error bars or significance tests.","section":"§4.3, Table 1"},{"comment":"The evaluation metric is circular in a mild but important way: TAP-Net is used both to generate the training trajectory labels (Section 3.4) and to track user-specified points in the generated videos during evaluation (Section 4.3). Because the model is trained to satisfy TAP-Net's tracking behavior, high Acc@0.05/0.01 may reflect TAP-Net's inductive biases rather than genuine point correspondence in the generated video. The authors should re-evaluate with an independent tracker (e.g., CoTracker or PIPs) or with human-annotated correspondences on a subset of the test set.","section":"§3.4, §4.3"},{"comment":"The paper claims better 'visual quality' but reports no image or video quality metric (e.g., FVD, CLIP score, or human preference). Table 1 only measures trajectory-following accuracy. A user study or standard quality metric is needed to support the visual-quality component of the central claim, which is currently asserted without quantitative evidence.","section":"§4.3"},{"comment":"The self-reported observation that purely planar horizontal shifts or zooms can leave content static shows that not all camera dynamics are faithfully encoded by point trajectories; the dolly-zoom requires a special combination of static subject and background zoom trajectories. This limits the claimed unification of camera and object motion and should be analyzed systematically, with a characterization of which camera motions are representable and quantitative camera-control evaluation (e.g., pose estimation) rather than only qualitative examples.","section":"§4.2, Figure 5"}],"minor_comments":[{"comment":"There are several typos: 'filted' should be 'filtered' in Section 3.4; 'we observer' should be 'we observe' in Section 4.2; and the team name in reference [24] is spelled 'Seawead' instead of 'Seaweed.'","section":"§3.4, §4.2, Reference [24]"},{"comment":"The definition of 'Appearance Rate' is vague ('whenever the user-specified trajectory is present'); please clarify whether this counts all frames in which the user trajectory exists and whether the tracker's visibility prediction is compared against user intent or against ground-truth visibility.","section":"§4.3"},{"comment":"The Gaussian width σ = 1/440 is given only in pixel units of the input image; since the latent grids of Seaweed-7B and Wan2.1-14B differ, please specify how σ scales across models and resolutions.","section":"§3.2"},{"comment":"The qualitative section does not state how representative the shown examples are or how failures were counted; a success-rate breakdown over the 100 test pairs, including the noted failure cases (rapid motion, object disassembly), would help calibrate the method's reliability.","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":"The core idea is reasonable, but the experimental section is not yet a fair test of the stated claims. If the authors cannot add non-circular, comparative evaluation, I would not support acceptance. I also note that several closely related trajectory-control methods (Motion Prompting, ReVideo, Trailblazer) are cited in the related work but never compared, which is a scope concern for a paper claiming state-of-the-art performance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: the method is real and the paper is worth a look, but the evaluation does not support its main claim. ATI injects point-trajectory features into a pretrained I2V transformer by bilinearly sampling a latent feature at each starting point and adding Gaussian-shaped masks at subsequent positions. That specific mechanism is new, and the tail-dropout regularizer is a sensible fix for the hallucinated-occluder problem when trajectories end early. Showing it on two different base models (Seaweed-7B and Wan2.1-14B) is useful evidence of compatibility.\n\nWhat the paper does well: the unified trajectory interface is genuinely simpler than separate camera/object/local modules, the write-up is clear about failure cases (rapid motion, disassembly, planar-only trajectories producing static content), and the qualitative examples appear compelling.\n\nThe soft spot is the quantitative section. Table 1 compares only ATI+Seaweed against ATI+Wan2.1 on 100 self-built image–trajectory pairs. There is no comparison with MotionCtrl, Motion Prompting, Perception-as-Control, VidCraft3, or any commercial tool, despite the abstract and conclusion claiming \"significantly better controllability and visual quality compared to prior approaches and commercial solutions.\" That claim is simply not tested. The metrics also reuse TAP-Net, the same tracker used to generate training labels, to measure trajectory following in generated videos. That creates a mild same-tool dependency; it does not invalidate the numbers by itself, but it does mean the accuracy figures may partly reflect TAP-Net's biases rather than solely the model's behavior. Combined with the small test set, this is weak evidence for the headline.\n\nThe limitation about planar trajectories is worth taking seriously: the method's core assumption is that all motion, including camera motion, can be encoded as 2D point trajectories. The dolly-zoom example works only because a static subject anchors the scene while background points zoom. It is unclear how well this generalizes to more complex 3D moves.\n\nOverall: the idea is plausible and the qualitative demonstration is strong, but the evaluation needs baselines and a tracker-independent metric before I would believe the superiority claim. The paper deserves a serious referee, but the revision should be substantial.\n\nFor peer review: I would send it out, with explicit instructions that the authors compare against at least one prior trajectory-control method and at least one commercial tool, and ideally add a human evaluation.\n\nBest,","headline":"ATI is a plausible trajectory-conditioning method with nice qualitative demos, but the quantitative evidence for its headline superiority claim is missing – no baselines, a small self-built test set, and a tracker reused from training labels.","tokens_in":9312,"tokens_out":2647,"would_cite":false,"duration_ms":26714,"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":"One trajectory input controls camera, object, and local motion in video generation.","keywords":["trajectory-based video generation","motion control","image-to-video generation","latent feature injection","diffusion transformer","camera control","object motion control","tail dropout regularization"],"falsifier":"Run the trained model on a first frame with strong depth layering — a near object in front of a distant wall — and specify a purely horizontal pan via parallel point trajectories across the whole image. If the output is a rigid 2D translation with no relative displacement between near and far content, while a physical camera pan would produce depth-dependent parallax, the claim that 2D trajectories alone encode camera dynamics is contradicted. A second decisive test: draw trajectories that force an object to split apart, and check whether the model follows the instruction or substitutes a natural alternative, as the paper reports it often does.","tokens_in":8353,"feed_emoji":"🎬","tokens_out":9131,"duration_ms":79845,"temperature":0.7,"pith_summary":"The paper sets out to prove that one input — trajectories of 2D points drawn on a single frame — can control every major kind of motion in generated video: camera movement such as pan and zoom, translation of whole objects, and fine-grained local deformation. Its wager is that these seemingly different motion types are all the same thing at the level of the diffusion model's latent space, and that a lightweight motion injector can translate point paths into latent instructions for a pretrained image-to-video model without task-specific modules. The authors build this injector into two large DiT backbones and report better controllability and quality than separate-module systems and commercial products, with one drawing interface for all motion. If the claim holds, the practical payoff is a single intuitive control surface for video generation and a control recipe that ports across backbone models. The paper itself flags where the equivalence strains: purely planar trajectory sets can leave content static and only produce 2D camera movement.","feed_headline":"Trajectories unify camera and object control for video AI","feed_subtitle":"Draw paths on one frame; a pretrained video model follows them for zooms, pans, and fine local motion.","key_machinery":"The load-bearing mechanism is the Gaussian motion injector. For each trajectory point it extracts a $C$-dimensional appearance feature $f$ from the first-frame latent by bilinear sampling at the start location $\\varphi_0$, then defines a spatial mask $P(f \\mid l_{i,j,t}) = \\exp(-\\|\\varphi_t - (i,j)\\|^2 / (2\\sigma))$ centered at the point's location $\\varphi_t$ in each later frame, with $\\sigma = 1/440$ so the weight falls to half at the nearest diagonal pixel. Blending these masked features into the latent stream before the DiT patchify layer gives the denoiser a per-timestep reading of \"this appearance belongs at this location now,\" converting sparse user strokes into dense soft guidance without changing the backbone architecture beyond fine-tuning all DiT parameters for around 50,000 iterations.","core_discovery":"ATI's central discovery is that a point trajectory becomes a latent instruction through two steps: sample the first frame's VAE feature at the point's starting location, then, in every later frame, re-deposit that feature over a small Gaussian neighborhood centered at the point's prescribed position. Each trajectory point thereby carries the appearance of the content it anchors, and the denoiser learns to move that content along the drawn path. The same mechanism covers local deformation, object translation, and camera dynamics, because camera motion is just many background points moving together. A tail-dropout regularizer — randomly truncating trajectories during training with probability 0.2 — teaches the model that a vanished point means \"no more instruction\" rather than \"something occludes it,\" which suppresses hallucinated occluders at early trajectory endings. On 100 hand-drawn trajectory sets, TAP-Net tracking shows points staying within 5% of the image diagonal on 55.9–59.0% of frames, with appearance rates of 65.5–67.9%, across the two backbones.","pith_inferences":["The injector couples only to the latent grid and the patchify input, so the same Gaussian feature-carrying scheme should transfer to any latent transformer video model, including text-to-video generators that synthesize a first frame; this is a testable extension the paper does not run.","The static-content failure on planar trajectories points to a concrete upgrade the paper leaves implicit: add a per-point depth or disparity channel so background trajectories scale with depth and produce genuine parallax, replacing the hand-placed static point its dolly-zoom needs.","Trajectory sets behave like a composition algebra — an affine transform applied to all points is a camera move, editing a subset is an object move — which an editor UI could expose as pan/zoom buttons that transform every point rather than as a separate camera module.","Because TAP-Net builds the training labels and also scores the outputs, the reported follow-rates partly measure tracker agreement with the generator's appearance changes; a human study or independent point-correspondence benchmark would separate control fidelity from tracking artifacts."],"forward_implications":["A single drawing interface can replace separate camera-control, object-drag, and local-deformation modules; users mix zooms, pans, and local edits in one set of strokes on the first frame.","The injector ports across DiT image-to-video backbones with a standard fine-tuning pass, so trajectory control becomes a general adapter rather than a bespoke model.","Trajectories that end before the final frame no longer invite hallucinated occluders, since tail dropout teaches the model that a vanished point simply stops being tracked.","Combining trajectories — radial background zoom plus a static subject point — reproduces compound effects like a dolly-zoom without task-specific code.","Extreme instructions (a point crossing half the frame width in two frames, or an object forced to split) are the documented failure boundary of the unified control."],"supporting_citations":[{"why":"Base image-to-video backbone (Seaweed-7B); ATI is fine-tuned on it and most qualitative results use it.","marker":"[24]"},{"why":"Second pretrained I2V backbone (Wan2.1-14B); demonstrates the injector transfers across models.","marker":"[28]"},{"why":"TAP-Net provides the point-trajectory training annotations and the tracking used to score instruction following.","marker":"[6]"},{"why":"DiT is the denoising architecture of both base models; the injector feeds the patchify layer.","marker":"[22]"},{"why":"MotionCtrl is the prior modular camera-plus-object controller that ATI claims to unify and outperform.","marker":"[31]"},{"why":"Motion Prompting is the closest trajectory-track conditioning approach the paper compares its unified latent injection against.","marker":"[8]"},{"why":"Training-free trajectory animation method used as the zero-shot baseline for comparison.","marker":"[39]"}],"fun_headline_variants":["Draw paths to steer video motion: camera, objects, and local effects","Unified trajectory control for video generation: one method, all motions","Trajectory-based video control: from camera pan to fine local motion","One trajectory framework for camera, object, and local motion in video","Control video motion with drawn paths: ATI unifies camera and object control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every motion a user wants, including camera dynamics, can be captured by the 2D movement of image-plane points with no explicit depth or 3D structure — a premise the paper itself strains when it notes that purely planar trajectory sets yield static content with only 2D camera movement.","fun_headline_variants_meta":{"raw":{"variants":["Draw paths to steer video motion: camera, objects, and local effects","Unified trajectory control for video generation: one method, all motions","Trajectory-based video control: from camera pan to fine local motion","One trajectory framework for camera, object, and local motion in video","Control video motion with drawn paths: ATI unifies camera and object control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000244,"raw_usage":{"total_tokens":1531,"prompt_tokens":943,"completion_tokens":588,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":493}},"tokens_in":559,"tokens_out":588,"duration_ms":5377,"temperature":1.0,"reasoning_tokens":493,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:55:50.985823+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained model on a first frame with strong depth layering — a near object in front of a distant wall — and specify a purely horizontal pan via parallel point trajectories across the whole image. If the output is a rigid 2D translation with no relative displacement between near and far content, while a physical camera pan would produce depth-dependent parallax, the claim that 2D trajectories alone encode camera dynamics is contradicted. A second decisive test: draw trajectories that force an object to split apart, and check whether the model follows the instruction or substitutes a natural alternative, as the paper reports it often does.","supporting_citations":[{"cited_title":"TAP-vid: A benchmark for track- ing any point in a video.Advances in Neural Information Processing Systems, 35:13610–13626, 2022","cited_arxiv_id":null,"evidence_quote":"TAP-Net provides the point-trajectory training annotations and the tracking used to score instruction following."},{"cited_title":"Motionctrl: A unified and flexible motion controller for video generation","cited_arxiv_id":null,"evidence_quote":"MotionCtrl is the prior modular camera-plus-object controller that ATI claims to unify and outperform."},{"cited_title":"Zero-shot controllable image-to-video animation via motion decomposition","cited_arxiv_id":null,"evidence_quote":"Training-free trajectory animation method used as the zero-shot baseline for comparison."}],"review_version":1}