{"id":"5137d740-5e3c-4ac3-b068-79e462967418","arxiv_id":"2412.05848","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A decoupled object/camera motion intensity estimator trained by contrastive ranking plus a diffusion transformer that injects the two scores to enable user-controllable video motion.","lead":"MotionStone trains a contrastive motion estimator to score object and camera motion separately in videos, then uses those scores to condition an image-to-video diffusion transformer so users can control motion strength. It reports improved generation quality and controllability over prior I2V models, though the evaluation relies partly on the estimator itself.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Absolute motion-intensity scores lack independent validation: the pseudo-label conversion is unspecified and Table 6 uses the estimator itself, so human-aligned 1-10 control is unsubstantiated.","rationale":"The paper's core proposal is a decoupled motion estimator trained with relative pairwise labels and regression pseudo-labels, plus a diffusion-transformer conditioning scheme. The ablations in Table 2 and the user study in Table 4 provide some support for the injection design and overall quality. However, the claim that the estimator yields stable, human-aligned absolute motion intensities is not supported by the evidence. The pseudo-label conversion in Sec. 3.3 is unspecified, so the regression targets that define the 1-10 scale are unknowable from the paper. Table 3 only shows relative ranking accuracy on the same video-pair distribution. Table 6, which reports a 2.52 motion strength error, evaluates generated videos using a motion estimator, presumably the same one used to condition the generator, so a low error can be achieved by the generator simply matching the estimator's biases rather than matching human perception. This circularity is the most load-bearing weakness because the user-facing intensity controls and the 'accurate estimation' claim depend on absolute calibration. The fix is straightforward: disclose the conversion, validate against independent human absolute ratings on external videos, and re-run Table 6 with a non-circular evaluator. Given the otherwise reasonable architecture and ablations, conditional acceptance remains appropriate.","tokens_in":16621,"tokens_out":7682,"duration_ms":77227,"concrete_test":"Re-run the Table 6 ablation using independent human annotators: generate videos at several input intensity levels (e.g., 2, 4, 6, 8) across diverse prompts, have annotators rate perceived object/camera motion intensity on the same 1-10 scale, and compute MSE between input values and mean human ratings. If the error is comparable to the reported 2.52, the control is real; if it is substantially larger, the low error arises because the generator matches the estimator's own biases rather than human perception.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that MotionStone's estimator provides stable, human-aligned object/camera motion intensities on a 1-10 scale rests on two unsupported links. First, Sec. 3.3 states that regression targets y_object and y_camera are 'approximate[d]' from tracking trajectories [49] combined with SAM masks [50], but the conversion from trajectories to intensity scores is never specified. This conversion defines the estimator's output range and the user-facing controls; without it, the absolute scale has no demonstrated semantics. Second, the only quantitative evidence for absolute accuracy is Table 6, which measures MSE between user-specified intensities and intensities 'obtained' by applying a motion estimator to generated videos. If that estimator is the same one used to train the generator, the loop can be closed by the generator learning the estimator's biases, so low error does not imply that human-perceived motion intensity matches the input. Table 3 only validates relative ranking on held-out video pairs from the same dataset, not absolute calibration or generalization to arbitrary in-the-wild videos. Thus the 'accurate' and 'controllable' claims are underdetermined.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MotionStone, an image-to-video (I2V) generation model built on CogVideoX that adds decoupled control of object and camera motion intensity. The authors train a motion estimator using pairwise human annotations of relative motion strength on 5,000 video pairs, together with a regression term whose pseudo-labels are derived from tracking and SAM-based masks. The resulting object and camera intensity scores are injected into the diffusion transformer through two separate MLPs concatenated into the time embedding. Experiments compare against I2VGen-XL, SVD, AnimateAnything, and CogVideoX using VBench metrics, user studies, ablations, and motion-strength error measurements, reporting improved quantitative scores and qualitative controllability.","tokens_in":16812,"tokens_out":4019,"duration_ms":39353,"significance":"If the claims hold, the decoupled motion estimator is a plausible scalable alternative to SSIM or optical-flow-based motion intensity for I2V training, and the injection scheme offers a practical way to expose object-versus-camera motion control to users. The paper's strengths include the relative annotation protocol (which is easier for annotators than absolute scoring), the compact TAdaConv-based estimator with dual heads, and ablations that consistently show each proposed module contributing. However, the absolute 1-10 scale is not independently validated, the regression pseudo-label construction is under-specified, and the control-precision evaluation in Table 6 is partly circular because it reuses the same estimator that provided the training signal.","major_comments":[{"comment":"The regression targets y_object and y_camera are said to be approximated from tracking trajectories [49] combined with SAM masks [50], but the conversion from trajectories to absolute intensity scores is never specified. This conversion defines the estimator's output scale and the user-facing 1-10 controls, so without it the absolute scale has no demonstrated semantics. Please specify the algorithm (e.g., how trajectory magnitudes are pooled, normalized, and mapped to the 1-10 range) and ideally validate the resulting pseudo-labels against human absolute ratings.","section":"Section 3.3, Eq. (7)"},{"comment":"The motion strength error is computed by applying a motion estimator to generated videos and comparing the result with the user-specified input intensity. If this is the same estimator used to produce training labels and to condition the generator during training, the low error (2.52) largely measures self-consistency between the generator and the estimator; the generator could learn the estimator's biases rather than human-perceived intensity. Please evaluate control precision with an independent human study (e.g., users judge which generated video matches a requested intensity) or with a separately-trained estimator not used in the pipeline.","section":"Table 6, Section F.1"},{"comment":"The annotation labels in Appendix B are -2, -1, 0, 1, 2, and the text says that significant differences are amplified by doubling the loss, but the main-text losses are simple max(0, s2 - s1) forms that assume one direction and no label-dependent weighting. As written, Eqs. (5) and (6) do not implement the described training procedure. Please provide the full loss that uses the signed labels and the amplification factor.","section":"Section 3.3, Eqs. (5)-(6) and Appendix B"},{"comment":"The headline improvements are small in absolute terms (e.g., Background Consistency 95.76% vs. 94.91% for CogVideoX-5B; Imaging Quality 62.29% vs. 61.99%). No error bars, confidence intervals, or significance tests are reported, and the evaluation set size is not stated. Without variance estimates, the state-of-the-art claim and the ablation conclusions are not statistically supported.","section":"Tables 1, 2, 5, 7"},{"comment":"Table 3 evaluates the motion estimator only on the validation split of the same human-annotated video-pair dataset used for training. There is no cross-dataset or in-the-wild evaluation, so the abstract's claim that the estimator is a 'general plug-in enhancer' for arbitrary videos is not supported. Please add a generalization study, e.g., zero-shot evaluation on another video dataset or on WebVID clips with independently collected relative judgments.","section":"Table 3 and Abstract"}],"minor_comments":[{"comment":"The row labeled 'MotionStone w/o S' appears to be the same as 'MotionStone w/ S' in Table 2 (both report 94.97% Background Consistency, 46.13% Aesthetic Quality, 60.73% Imaging Quality). The label should be 'MotionStone w/ S' or the contents should be corrected to avoid confusion.","section":"Table 5, Row 2"},{"comment":"The evaluation set is referred to as 'WebVID validation set [1]' in one place and 'WebVID [22]' in another, but reference [22] is WebVision, not WebVid. Please use the correct citation consistently.","section":"Section 4.1 and References [1], [22]"},{"comment":"The amplification rule for significant differences is described only in prose. Even after the main-text loss is corrected, please give the exact formula for how the magnitude labels (1 vs. 2) enter the loss.","section":"Appendix B"},{"comment":"The 100,000 training videos are described only as 'high-quality videos collected by ourselves'; please provide details on curation, filtering criteria, licenses, and whether there is any overlap with the WebVID evaluation set.","section":"Section 4.1"},{"comment":"Several qualitative claims about camera pan/zoom and object motion are supported by still frames; since these effects are temporal, please point readers explicitly to the supplemental video files rather than relying on static snapshots for verification.","section":"Figures 4-9"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid engineering paper with a genuinely new annotation scheme—pairwise human ranking of object versus camera motion—and an injection design that improves controllable I2V generation. The relative motion estimates are likely fine; the absolute 1-10 intensity scale is not validated.\n\nThe core idea is simple and plausible. Instead of asking annotators to assign absolute motion scores, they compare pairs of videos and say which has stronger object motion and which has stronger camera motion. They train a TAdaConv backbone with two MLP heads using a ranking loss plus a regression loss on pseudo-labels derived from tracking trajectories. That regression step is where the paper gets murky: the conversion from trajectories to a 1-10 score is never specified, and that conversion defines the entire output range and the user-facing controls. Without it, there is no demonstrated semantics for any particular number on the scale.\n\nThe ablation study is honest and supports the main modules. Removing the estimator or the decoupled injection hurts quality. Replacing the estimator with SSIM or feature difference also hurts, which makes sense because those metrics do not disentangle object and camera motion. The gains over CogVideoX are small but consistent across metrics, and the user study supports the claims of better text and motion fidelity. I would have liked error bars on the VBench numbers—a 0.85% background consistency difference is hard to interpret without variance.\n\nTable 6 is the weakest spot. They measure motion strength error by generating videos with a specified intensity, then applying a motion estimator to the output and computing MSE. If that estimator is the same one used to train the generator, the low error mostly shows self-consistency, not that the generated intensity matches what a human would perceive. The stress-test note is right about this. It is fixable: use an independent estimator, or get human absolute ratings on a 1-10 scale, or at least show that the estimator's absolute scores correlate with held-out human rankings on unseen videos.\n\nThe citation pattern looks fine and the code/data release would answer most of the remaining questions. This is a useful paper for anyone working on motion control in video diffusion models, and it deserves a serious referee. I would recommend conditional acceptance with a request to clarify the pseudo-label conversion and add a non-circular control evaluation.","headline":"A useful decoupled motion-intensity estimator with a novel ranking-based annotation scheme, but the absolute 1-10 scale is underspecified and the control evaluation is circular.","tokens_in":17360,"tokens_out":2242,"would_cite":true,"duration_ms":21388,"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":"MotionStone claims that decoupling object and camera motion intensity, with scores learned from pairwise human comparisons and tracking-based pseudo-labels, gives image-to-video diffusion models stable, human-aligned control over motion.","keywords":["image-to-video generation","motion intensity estimation","object motion","camera motion","contrastive learning","diffusion transformer","motion control","decoupled conditioning"],"falsifier":"Render or record clips with calibrated motion—a camera panning at known speeds and an object moving at known velocities—then run the estimator on them and check whether its object and camera scores are monotone in the true speeds and whether pairwise rankings match the ground truth; non-monotone or inverted scores on these calibrated cases would falsify the claim that the estimator aligns with human perception.","tokens_in":16431,"feed_emoji":"🎬","tokens_out":7786,"duration_ms":66237,"temperature":0.7,"pith_summary":"MotionStone is built on the idea that motion intensity for image-to-video generation should be two separate numbers, one for object motion and one for camera motion, rather than one global value. Because people find it hard to label absolute intensity, the paper trains a motion estimator with simple pairwise comparisons—which of two clips moves more—and then anchors its 1-to-10 scale with tracking-derived pseudo-labels. The frozen estimator's two scores are injected into a pretrained diffusion transformer through separate embeddings added to the time-step condition, so the model can tell camera control from object control. The paper reports that this decoupled scheme improves temporal consistency and image quality over fixed-intensity, SSIM-based, and feature-difference baselines, and that users can independently dial object and camera intensity. The central claim is that a reliable, human-aligned decoupled motion estimator is what makes fine-grained motion control in image-to-video generation work.","feed_headline":"Splitting motion into object and camera scores sharpens video control","feed_subtitle":"A contrastive estimator feeds a diffusion transformer, letting users dial object and camera motion from 1 to 10.","key_machinery":"The load-bearing mechanism is the decoupled motion estimator plus its injection scheme. The estimator uses a TAdaConv backbone to extract spatiotemporal motion features, then two MLP heads predict object motion and camera motion scores in the 1-to-10 range. It is trained with a pairwise ranking loss on relative human annotations and a regression loss whose pseudo-labels come from tracking trajectories combined with segmentation masks. In the generator, the two scores pass through separate MLPs, are concatenated, and are added to the time embedding of a diffusion transformer, so each control signal is kept distinct during denoising.","core_discovery":"The paper's central discovery is that a lightweight motion estimator can decompose the motion in a raw video clip into an object score and a camera score, and that these decoupled scores are the right conditioning signal for a video diffusion model. The estimator is trained in two complementary ways: contrastive ranking on 5,000 video pairs labeled by relative human judgment, and regression on pseudo-labels computed from tracking trajectories combined with segmentation masks. During generation training the estimator is frozen, and its two outputs are mapped through separate MLPs, concatenated, and added to the diffusion time embedding, which keeps the two control channels from blurring together. On the reported benchmarks MotionStone improves background consistency, aesthetic quality, and imaging quality over the compared image-to-video models, and the estimator itself predicts human-annotated relative motion in video pairs with 72.80% accuracy versus 44.56% for an SSIM-based proxy.","pith_inferences":["The same contrastive setup could be extended to other subjective video properties, such as motion naturalness or speed categories, without changing the annotation protocol, since it only requires relative labels.","A concrete offline use of the plug-in framing would be corpus filtering: the estimator could score large unlabeled video collections and curate training sets with a desired balance of object and camera motion.","The paper validates relative ranking and generation quality but does not compare its 1-to-10 scores against human absolute ratings; a direct calibration study would be the quickest way to test whether the estimator's scale is truly human-aligned.","The decoupled injection design suggests that other paired or opposing conditions, such as content motion versus style motion, could benefit from separate embedding channels added to the time-step modulation rather than a single merged condition."],"forward_implications":["Users can set object motion and camera motion independently on a 1-to-10 scale, and the generated video's movement follows the requested levels.","Training with estimated, per-video motion intensities yields better background consistency and frame quality than training with a fixed default intensity.","Replacing whole-video motion proxies such as inter-frame SSIM or feature difference with the decoupled estimator improves both generation quality and motion controllability.","Because the estimator is frozen and only supplies additional conditioning, it can be attached to an existing diffusion-transformer video generator as a plug-in enhancer.","On the validation video-pair set, the estimator predicts human-annotated relative motion with 72.80% accuracy, well above an SSIM-based proxy's 44.56%."],"supporting_citations":[{"why":"Supplies the TAdaConv backbone that extracts spatiotemporal motion features for the estimator.","marker":"[19]"},{"why":"Provides the pairwise ranking loss used to train the estimator from relative human annotations.","marker":"[25]"},{"why":"Supplies tracking trajectories from which the object and camera motion pseudo-labels are computed.","marker":"[49]"},{"why":"Supplies the object masks used with tracking to separate object motion from camera motion in pseudo-label generation.","marker":"[50]"},{"why":"Provides the pretrained diffusion transformer video generator that MotionStone builds on and compares against.","marker":"[52]"},{"why":"Provides the inter-frame SSIM motion-intensity baseline used in ablations and comparisons.","marker":"[8]"},{"why":"Provides the feature-difference motion-intensity baseline used in ablations and as a comparison method.","marker":"[11]"},{"why":"Supplies the evaluation metrics used for quantitative comparisons of generated videos.","marker":"[20]"}],"fun_headline_variants":["Decoupling motion into object and camera boosts video realism","Separate control of object and camera motion in I2V","Contrastive motion scores separate camera from object movement","MotionStone: dial object motion and camera motion independently","Object vs camera motion: a new lever for video generation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the assumption that the tracking-and-segmentation pseudo-labels, which fix the estimator's 1-to-10 scale, are faithful proxies for what people mean by object and camera motion intensity.","fun_headline_variants_meta":{"raw":{"variants":["Decoupling motion into object and camera boosts video realism","Separate control of object and camera motion in I2V","Contrastive motion scores separate camera from object movement","MotionStone: dial object motion and camera motion independently","Object vs camera motion: a new lever for video generation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000471,"raw_usage":{"total_tokens":2353,"prompt_tokens":968,"completion_tokens":1385,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":1306}},"tokens_in":584,"tokens_out":1385,"duration_ms":9365,"temperature":1.0,"reasoning_tokens":1306,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:16:21.889505+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render or record clips with calibrated motion—a camera panning at known speeds and an object moving at known velocities—then run the estimator on them and check whether its object and camera scores are monotone in the true speeds and whether pairwise rankings match the ground truth; non-monotone or inverted scores on these calibrated cases would falsify the claim that the estimator aligns with human perception.","supporting_citations":[{"cited_title":"Rankiqa: Learning from rankings for no-reference image quality assessment","cited_arxiv_id":null,"evidence_quote":"Provides the pairwise ranking loss used to train the estimator from relative human annotations."},{"cited_title":"Spatialtracker: Tracking any 2d pixels in 3d space","cited_arxiv_id":null,"evidence_quote":"Supplies tracking trajectories from which the object and camera motion pseudo-labels are computed."},{"cited_title":"Livephoto: Real image animation with text-guided motion control","cited_arxiv_id":null,"evidence_quote":"Provides the inter-frame SSIM motion-intensity baseline used in ablations and comparisons."},{"cited_title":"Animateanything: Fine-grained open domain image animation with motion guidance","cited_arxiv_id":null,"evidence_quote":"Provides the feature-difference motion-intensity baseline used in ablations and as a comparison method."},{"cited_title":"Vbench: Comprehensive bench- mark suite for video generative models","cited_arxiv_id":null,"evidence_quote":"Supplies the evaluation metrics used for quantitative comparisons of generated videos."}],"review_version":1}