acceptodds
Under review as a conference paper at ICLR 2027

DragPhysSVG: Bridging Vector Graphics and Particle Dynamics for User-Controlled Animation

Abstract

Scalable vector graphics (SVGs) represent visual content as editable vector paths, but their path structure reflects how an image is drawn rather than how the depicted object is physically organized. This mismatch makes physically grounded animation of unrigged SVGs challenging: parts belonging to the same body should move coherently, articulated parts should remain connected, and embedded details should follow their supporting bodies. Existing SVG animation methods often rely on predefined structure or direct path deformation, while particle-based physical models do not preserve the path organization required for editable vector output. We introduce DragPhysSVG, a structured particle representation that bridges vector graphics and particle-based dynamics. Specifically, SVG paths are sampled into particles with particle-to-path correspondence, while a vision-language model infers their latent physical organization, including body groupings, embedded details, and articulated connections. These relations form a structured graph augmenting the particle representation with the original SVG organization. Given a user drag, a pretrained physical diffusion model predicts particle motion, and our articulated-flow guidance preserves attachments and constrains relative motion across connected parts. Resulting trajectories are transferred back to the original paths through path-local deformation, producing editable animated SVGs. Experiments on articulated SVGs demonstrate improved structural coherence, joint stability, and drag controllability while preserving vector organization. Video animations are available on https://dragphyssvg.github.io/dragphyssvg/

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Controllable SVG Animation, Vector Graphics Animation, Physics-Grounded Animation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_jJLCOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes DragPhysSVG, a framework for drag-controlled physical animation of unrigged SVGs. The pipeline consists of: (1) Inferencing physical organization (body, path, joint, stiffness, and mapping between path and particles), BODY-AWARE SVG-TO-PARTICLES REPRESENTATION, from SVG paths via VLM; (2) Predicting physical aware particle deformation via a frozen diffusion model (PhysCtrl); (3) three post-processing stages—Articulated-Flow Guidance (Stage 1), Stiffness-Aware Projection (Stage 2), and Vector Path Deformation (Stage 3)—that transfer particle trajectories back to editable SVG paths. Experiments compare against physical-video baselines (PhysCtrl, Force Prompting) and one SVG animation baseline (AniClipart).

Critical strengths
  1. Well-motivated problem. Physically grounded animation of unrigged SVGs is a genuinely under-explored problem. The mismatch between SVG path structure (drawing order) and physical organization (semantic bodies and joints) is real and well articulated.

  2. Sensible modular design. The decomposition into VLM-based structure inference, particle-based physical simulation, and structure-guided motion approximation and transfer is clean and easy to follow. The VLM inference and the frozen physical diffusion model are effective to solve two challenges of the whole system. The rest challenge is solved by the structure-guided motion approximation and transfer module, which follows an intuitive and easy way by operating on each body and joint.

Critical weaknesses

W1. Scope over-claiming: the method only handles articulated deformation The title and framing ("STRUCTURED PARTICLE DYNAMICS FOR SVG ANIMATION") suggest a general-purpose SVG animation framework. It is better to restrict its claims to articulated SVG animation.

W2. Only one SVG animation baseline is compared The paper compares only against AniClipart as an SVG animation baseline. This is insufficient given recent work on non-rig SVG animation, including:

  • Vector Prism, cvpr26
  • PhysAnimator: Physics-Guided Generative Cartoon Animation (CVPR 2025)
  • LiveSVG: Zero-Shot SVG Animation via Video Generation (2026)
  • VAnim: Rendering-Aware Sparse State Modeling for Structure-Preserving Vector Animation, ICML26 The authors should address: Representation-level differences, PhysCtrl advantage, VLM reliability by comparing with above 4 methods. Without such analysis, the reader cannot judge the technique contribution of the proposed pipeline and the 3 post processing stages. The only conclusion can we draw is that it is novel and workable system for physical animation of unrigged SVG.

W3. Computational cost and real-time feasibility are under-reported The appendix contains some computation cost information, but it is far from sufficient to form a complete picture of runtime & cost of the system. The pipeline involves a VLM inference pass, particle sampling, a frozen physical diffusion model (which itself requires iterative denoising), and three post-processing stages. For a user-controllable, drag-based system, latency is not a secondary concern—it is central to usability.

Questions for authors

Error propagation: How do VLM errors, PhysCtrl errors, and motion-transfer errors propagate to the final animation? Please provide at least one illustrative example for each.

Other questions can be found in "Critical weaknesses".

4 comments
AuthorsOpenReview ↗

Official Comment: Response to Reviewer jJLC

Thank you for the insightful and constructive feedback. We address each concern below.

W1. Scope and supported animation types

We appreciate the suggestion to clarify the scope. Our method supports both joint articulation and within-body deformation according to body rigidity, so describing it solely as articulated animation would omit its support for soft bodies and mixed-rigidity structures. More precisely, it targets force-driven planar geometric animation of single-object, unrigged SVGs, covering soft bodies, articulated rigid bodies, and mixed-rigidity structures. Attribute changes, path appearance or disappearance, and out-of-plane 3D rotations are outside the current formulation. We will revise the Introduction to state these capabilities and boundaries explicitly.

W2. Additional baselines and evidence for the technical contributions

We have added a comparison with Vector Prism and updated Figure 4, Table 1, and Table 2. As public code is not yet available for PhysAnimator, LiveSVG, and VAnim, we discuss their methodological differences below.

  • PhysAnimator performs image-space deformable-mesh simulation and uses the resulting warped sketches to guide video generation. It supports forces and rigging points, but its output is raster animation. Our representation retains explicit particle-to-path correspondences to produce editable SVG animation.
  • LiveSVG fits SVG geometry to a generated target video using group-level homographies and path-level Bézier offsets. Its motion is driven by the reference video, whereas our representation explicitly encodes bodies, attachments, and joints to guide how drag-induced motion propagates across components.
  • VAnim learns text-conditioned sparse state updates over a persistent SVG DOM with rendering-aware reinforcement learning. Preserving SVG identities and topology differs from modeling physical attachments and articulated force responses, which our structured particle representation explicitly supports.

Evidence for the components. Articulated-flow guidance operates during denoising, so the three stages are not all post-processing operations. Table 3 compares guidance disabled with the default setting while retaining projection and transfer, showing reduced joint drift on both datasets. The existing qualitative ablations illustrate the effects of removing stiffness awareness and inter-part relations, while our additional manual structure inspection and VLM replacement experiment address inference reliability. These experiments provide complementary evidence, with quantitative isolation currently strongest for guidance.

PhysCtrl supplies force-conditioned particle trajectories that can be associated with SVG paths and adjusted during sampling without retraining. The proposed representation supplies the body organization and path correspondence needed to turn these trajectories into structured, editable SVG motion. Our experiments evaluate this combination. PhysCtrl is one motion prior used in our experiments, and our framework is in principle extensible to other motion priors.

Added Vector Prism comparison. Vector Prism uses VLM-based semantic grouping and animation planning followed by CSS animation generation. Our added comparison shows cases with disrupted connections between parts and motion misaligned with the applied force, while our method better preserves attachments and follows the force direction (Figure 4). Tables 1 and 2 now include Vector Prism in the quantitative and VLM-based evaluations. Vector Prism obtains higher subject consistency and lower all-body distortion on dataset (a). The examples in Figure 4 show limited motion, which illustrates why these scores alone do not establish force responsiveness, as lower all-body distortion can reflect less deformation rather than more realistic dynamics. Our method maintains high subject consistency while producing force-responsive motion.

AuthorsOpenReview ↗

Official Comment: Response to Reviewer jJLC (continued)

W3. Preprocessing cost and per-drag latency

We profiled 20 SVGs with 3 to 51 paths, separating preprocessing from repeated drag computation. As specified in Appendix A.1, Qwen structure inference uses two NVIDIA L40S GPUs, while motion generation uses one L40S. The timings are summarized below.

StageFrequencyTime: mean (range)
VLM structure inferenceOnce per SVG~1 min per request, ~4 min per SVG (4 samples)
Particle sampling + relation graphOnce per SVG5.1 s (1.3 to 24.4 s)
Guided PhysCtrl sampling (25 steps)Per drag4.0 s (3.6 to 6.6 s) (2.8 s without guidance)
Stiffness projection + SVG outputPer drag0.12 s (0.05 to 0.46 s)

Each drag therefore needs 4.1 s of computation on average (3.7 to 6.7 s). The output is a native SVG animation, so playback requires no further model inference. The ~15 s reported in the appendix was measured from end to end, including per-SVG particle sampling, model loading, and rendering a GIF preview. Most of the drag time comes from the denoising steps of the frozen generator, with our guidance and projection together adding about 1.3 s. Although our system does not provide fully real-time end-to-end feedback, it supports repeated drag interactions with rapid animation previews. We have updated Appendix A.1 with this timing breakdown and distinguish preprocessing, per-drag computation, and preview rendering.

Q1. Error propagation through the full pipeline

Appendix A.4 provides one example for each stage, showing how intermediate errors affect the final animation with the full pipeline active.

  • (a) VLM error: On the snowman, the inferred structure omits the joint between the hat and head. Guidance and projection therefore lack the attachment constraint between these parts, so the hat drifts away and intersects the head under gravity. Later stages cannot reliably recover a connection absent from the relation graph.
  • (b) PhysCtrl error: On the balloon dog, applying five times the default force produces local distortions in the particle trajectories. Stiffness projection maps each body to a valid rigid transform plus a bounded residual, preserving shape and connectivity, but correcting geometry does not guarantee that the underlying motion is physically correct.
  • (c) Motion-transfer error: On the balloon, the soft string’s motion is represented by particle residuals and interpolated onto its path. Rapid deformation over a few frames moves neighboring control points inconsistently, causing the thin path to fold over itself.

These cases motivate stronger VLM structural reasoning, physical-validity constraints on generated trajectories, and curve-aware motion transfer.

Thank you again for your encouraging feedback and support. We hope our clarifications, along with the added Vector Prism comparison, runtime breakdown, and error-propagation analysis, address your concerns.

Reviewer_jJLCOpenReview ↗

keep my rating

Thank you for the thorough rebuttal and additional experiments. They addressed my concerns. I support acceptance.

Thank you for the positive feedback and for supporting acceptance. We appreciate your confirmation that our rebuttal and additional experiments addressed your concerns.

Reviewer_98ypOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DragPhysSVG for drag-driven animation of static, unrigged SVGs, with the output kept as editable vector paths. A VLM infers how the paths form physical bodies, which paths are only surface details, the stiffness of each body, and the joints between bodies. This structure is attached to a particle representation of the SVG. A frozen PhysCtrl model generates the particle motion with articulated-flow guidance. A stiffness-aware projection then enforces the inferred structure, and the result is mapped back to the original paths. Experiments on 142 SVGs show lower rigid-body distortion and joint drift than the baselines (PhysCtrl, Force Prompting, AniClipart), while keeping good force alignment.

Critical strengths
  • The main strength for me is the problem setting and the representation. The paper makes a good point that SVG paths follow how an image is drawn rather than how the object is physically built. The proposed representation keeps bodies, joints and path correspondences together, so the motion from a particle-based physics model can be turned back into an editable SVG. I think this is useful in practice.

  • The method is also training-free and fairly modular, and the qualitative results in Fig. 4 look clearly better to me. The kite frame and the crab’s body stay rigid and the claws rotate around their joints, while PhysCtrl shears the claw and distorts the kite outline.

Critical weaknesses
  • The effect of each component is mainly shown qualitatively. The comparison with existing methods in Table 1 is informative, but the role of each component, especially the articulated-flow guidance, is only shown through a few examples (Figs. 5 and 6). Since the guidance is one of the main technical contributions, a small quantitative ablation (e.g., λ\lambda = 0 vs. the default setting, with the metrics in Table 1) would make it easier to see how much it adds on top of the stiffness-aware projection.

  • The dependence on the VLM is not discussed much. The whole pipeline relies on the inferred bodies, stiffness levels and joints. Apart from a brief remark in the ethics statement, I did not find a discussion of failure cases, for example what the animation looks like when a joint or a rigidity label is wrong, and how often this happens. Relatedly, it was not clear to me where the labels used in the structure metrics (rigid-body distortion and joint drift) come from. If they come from the VLM, it would be good to state this, since these metrics would then partly reflect agreement with the inferred structure.

Questions for authors
  • Could you add a quantitative version of the ablation in Sec. 4.3, at least for the guidance (λ\lambda = 0 vs. the default), using the metrics in Table 1?

  • Where do the labels used in the structure metrics come from? How does the method behave when the VLM infers a wrong structure, such as a missed joint or a wrong rigidity label?

4 comments
AuthorsOpenReview ↗

Official Comment: Response to Reviewer 98yp

Thank you for the helpful and constructive comments. We respond to each concern below.

W1. Quantitative contribution of articulated-flow guidance

We provide a quantitative ablation comparing articulated-flow guidance disabled lambda=0 with the default setting lambda=0.25 on both datasets. The disabled-guidance variant retains stiffness-aware projection and motion transfer. Joint drift decreases from 0.0078 to 0.0064 on AniClipart (17.9%) and from 0.0055 to 0.0050 on SVG-Bench (9.1%). Force alignment improves modestly, while rigid-body distortion on AniClipart and subject consistency on SVG-Bench are unchanged at the reported precision. These results support the contribution to force following and structure preservationan, with additional contribution to joint consistency. The quantitative ablation is included in Table 3.

DatasetlambdaA_force↑E_body_rigid ↓E_body ↓E_joint ↓BNE0 ↓SC ↑
AniClipart00.9560.01020.02090.00780.3780.977
AniClipart0.250.9590.01020.01940.00640.3620.978
SVG-Bench00.9680.00580.01020.00550.2210.987
SVG-Bench0.250.9750.00570.01000.00500.2200.987

W2.1. VLM dependence: Reliability of inferred structures and consequences of errors

We repeated the same structure-inference procedure with GPT-5, obtaining four responses per SVG as in the original pipeline. GPT-5 produces more consistent outputs, with self-consistency of 0.73, compared with 0.47 for Qwen. The table below reports downstream animation results over all samples from both datasets evaluated in the paper, showing similar overall performance with GPT-5 and Qwen as the structure-inference VLM. The disagreements primarily concern the granularity of body partitioning rather than conflicting interpretations of the underlying physical structure. Our original pipeline addresses this variability by voting among responses that pass the structural validator and connect only touching paths by joints, the response whose body partition best agrees with the other responses. The Qwen (Final) column shows the benefit of voting: compared with individual Qwen candidates, the selected outputs receive more reasonable judgments for body partitioning (92% vs. 85%), inter-body relations (80% vs. 67%), and rigidity assignments (88% vs. 80%).

ComponentGPT-5 candidates judged reasonableQwen candidates judged reasonableQwen (final) reasonable
Body partition91%85%23/25 (92%)
Inter-body relations76%67%20/25 (80%)
Rigidity assignments89%80%22/25 (88%)

Although the final structures are more often judged reasonable than the raw Qwen candidates in this sample, connection errors remain. Appendix A.4 illustrates a missed joint between the snowman's hat and head. Guidance and projection then lack the attachment constraint needed to keep the parts connected, allowing the hat to drift and intersect the head.

As a separate sensitivity analysis, we replaced Qwen with GPT-5 and generated animations over both evaluation datasets. The table below shows similar overall performance with GPT-5 and Qwen as the structure-inference VLM.

Structure-inference VLMA_force ↑BNE0 ↓SC ↑
GPT-50.9690.2460.9840
Qwen0.9700.2620.9844

Overall, downstream force-control accuracy and video quality remain largely unchanged, with modest gains in structure preservation benefiting from GPT-5’s more stable inference and semantic decomposition. Annotation differences mainly concern granularity, suggesting that our framework is not tied to a particular VLM, while stronger VLMs can improve inference stability and motion quality.

AuthorsOpenReview ↗

Official Comment: Response to Reviewer 98yp (continued)

W2.2. Evaluation labels and sensitivity to their source

The labels come from the same Qwen3-VL-32B structure annotations that our method receives as input. Joint anchor locations are computed geometrically from contact between SVG paths, and pairwise-distance preservation within a body and anchor drift across a joint are measured on CoTracker3 trajectories.

These metrics quantify how well the generated motion preserves the specified body and joint structure. To analyze the sensitivity to the annotation source, we have therefore re-evaluated all methods’ existing outputs using GPT-5 structure annotations. The tables below compare results using Qwen and GPT-5 annotations as references, keeping all generated animations fixed and changing only the evaluation annotations. On SVG-Bench, our method achieves the lowest errors on all three metrics under both annotation sources. On AniClipart, it retains the lowest joint drift, while the rigid-body ranking changes: Vector Prism obtains 0.0155 under GPT annotations, compared with 0.0166 for our method. Thus, the joint-consistency advantage persists, while some rigid-body comparisons are more sensitive to the annotation source. Overall, the comparative results remain consistent, supporting the stability of our conclusions across annotation sources.

AniClipart

MethodE_body_rigid ↓ (Qwen ref.)E_body_rigid ↓ (GPT ref.)E_body ↓ (Qwen ref.)E_body ↓ (GPT ref.)E_joint ↓ (Qwen ref.)E_joint ↓ (GPT ref.)
PhysCtrl0.02480.07710.02470.03390.00960.0087
Force Prompting0.02070.02130.02440.02430.01290.0086
AniClipart0.02870.01910.02320.02060.00890.0085
Vector Prism0.01330.01550.01420.01510.01390.0119
Ours0.01020.01660.01940.01850.00640.0058

SVG-Bench

MethodE_body_rigid ↓ (Qwen ref.)E_body_rigid ↓ (GPT ref.)E_body ↓ (Qwen ref.)E_body ↓ (GPT ref.)E_joint ↓ (Qwen ref.)E_joint ↓ (GPT ref.)
PhysCtrl0.02380.02160.02330.02270.01110.0098
Force Prompting0.01510.01390.01620.01620.00790.0086
Vector Prism0.01680.01710.01570.01540.00730.0062
Ours0.00570.00870.01000.01040.00500.0036

Thank you again for your positive assessment and constructive feedback. We hope the quantitative ablation, VLM-robustness analysis, and cross-annotation evaluation resolve your remaining concerns.

Reviewer_98ypOpenReview ↗

Response to Authors

Thank you for the thorough rebuttal and additional experiments. The quantitative ablation, cross-VLM sensitivity analysis with GPT-5, and failure-case discussion address my concerns well. Since my initial rating was already positive, I will maintain my original score and support acceptance.

Thank you for your positive assessment and for recognizing that our rebuttal and additional experiments address your concerns. We appreciate your continued support for acceptance.

Reviewer_4N5fOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to animate a static SVG by dragging its parts, without first building a skeleton. The challenge is that SVG paths describe how an image is drawn but don't indicate which parts should move together. In the proposed method, a VLM identifies the parts, their stiffness, how they connect, and which details belong to each part. The SVG is then converted into particles, keeping this structure and links to the original paths. A frozen PhysCtrl model generates their motion, while articulated-flow guidance steers its predictions toward joint-following motion.

Critical strengths

The overall idea is well aligned with its core design. The proposed test-time pipeline further extends PhysCtrl's ability to handle SVG dynamics, improving force prompt following and structure maintenance.

Critical weaknesses
  1. The main pipeline and evaluation of this paper emphasize the ability to maintain rigid parts and structure (evaluation on Ebody), while the prompt for PhysCtrl and Force Prompting's video generator emphasizes a soft, elastic response. This may create an unfair comparison.
  2. Compared with AniClipart, which generates text-driven dynamics, and physical-video baselines designed for diverse real-world images, this paper focuses on a narrower task of local force-driven SVG animation. This makes interaction diversity particularly important for demonstrating the task’s value. However, the evaluation tests only four inward force directions at a fixed magnitude. This limited range further narrows the significance of the demonstrated application.
  3. The evaluation of the same objects with different materials is not, at least qualitatively, evaluated, which may be another limitation.
Questions for authors
  1. As PhysCtrl produces motions of 3D particles, how to make it work for the movement of 2D particles? More details would be appreciated.
  2. For material behavior, Table 2 reports a VLM win rate of only 51.2% against PhysCtrl, so the improvement is not clearly established. I believe a user study would be necessary.
  3. Could the author show some failure cases of the proposed methods using the full setup?
2 comments
AuthorsOpenReview ↗

Official Comment: Response to Reviewer 4N5f

Thank you for the thoughtful feedback and constructive suggestions. We address each concern below.

W1. Baseline prompting and fairness of structural evaluation

We first clarify that both physical-video baselines use consistent elastic-material descriptions. PhysCtrl and Force Prompting share the same core prompt, including "deforms softly like rubber and springs back." PhysCtrl additionally specifies explicit elastic parameters (log E = 4, ν = 0.4), while Force Prompting, which lacks material parameters, uses an extended description of the same elastic behavior. Importantly, our method uses the same frozen elastic PhysCtrl checkpoint and identical physical parameters as PhysCtrl. Therefore, our improvements do not arise from using a more rigid underlying physics model.

Our goal is not to maximize rigidity, but to preserve component-specific physical behavior. Elastic objects can contain components with different stiffness levels. Our structured representation explicitly distinguishes rigid, firm, and soft bodies, allowing rigid parts to retain their shapes, flexible parts to deform, and connected parts to articulate. As shown in Fig. 4, the kite's frame remains intact while its streamers bend, and the crab's claws rotate around their joints rather than collapsing. This demonstrates material-dependent deformation rather than globally enforced rigidity.

E_body is also not our sole evaluation criterion. Table 1 additionally reports force alignment, joint drift, topology preservation, and subject consistency, while Table 2 evaluates perceptual material behavior and physical plausibility. Compared with PhysCtrl, our force alignment improves from 0.925 to 0.959 and 0.859 to 0.975 across the two datasets, while joint drift decreases from 0.0096 to 0.0064 and 0.0111 to 0.0050, respectively. These results show that improved structural preservation is achieved alongside stronger force controllability, rather than simply suppressing motion.

We will clarify the baseline prompt settings and evaluation rationale in the revision. Our contribution is to enable structurally coherent, material-dependent articulation under an elastic physics prior, rather than indiscriminately reducing deformation.

W2. Diversity of force directions and magnitudes

We use four inward force directions at a fixed magnitude in order to provide controlled, matched conditions for quantitative comparison with prompt-based methods; this protocol does not represent the full range of interactions supported by our method. Section 01 of the supplementary page provides qualitative examples with varied force directions and four distinct force magnitudes, demonstrating that our method supports diverse directions and tolerates a broader range of force magnitudes. This flexibility illustrates the practical value of drag-based SVG animation by letting users explore different motions simply by varying how they drag.

W3. Material variations on the same SVG

We have added comparisons of the same SVGs under elastic, plasticine, sand, and rigid material settings in Appendix A.3 and on the supplementary page. All four settings use gravity and ground collision. The released PhysCtrl's non-elastic materials (plasticine, sand and rigid) can only be driven along the gravity direction, which is why the original arbitrary-direction drag comparison used the elastic model.

The examples cover a single soft body, articulated rigid bodies, and mixed-rigidity structures. They exhibit rebound, permanent deformation, spreading, and rigid motion under the respective material settings. The material-dependent dynamics come from PhysCtrl, while our representation and transfer impose the inferred body organization and attachments. These qualitative examples demonstrate compatibility with multiple material priors.

Q1. Adapting 3D particle trajectories to 2D SVGs

We use PhysCtrl unchanged as a frozen 3D motion prior. We embed the sampled 2D particles into the image plane by setting z=0z=0, with the drag point and force also confined to this plane. PhysCtrl predicts 3D trajectories, from which we retain only the xx and yy coordinates. Our structure-guided motion processing and SVG deformation then operate entirely in 2D, requiring no retraining of PhysCtrl. The resulting motion is a planar projection of a 3D prior rather than a dedicated 2D physical simulation. We will add this implementation detail and its scope to the manuscript.

AuthorsOpenReview ↗

Official Comment: Response to Reviewer 4N5f (continued)

Q2. Material-behavior evidence and human evaluation

To complement the VLM-based evaluation, we conducted a user study with 30 participants on 13 cases, assessing force-direction following, physical plausibility, and subject consistency. For each case, participants were shown the original image and force arrow and selected the best animation separately for each criterion. Physical plausibility was assessed considering the object's structure and material properties, while subject consistency was assessed relative to the original image.

The table below reports each method’s selection-based win rate. Ours achieves the highest win rate across all three criteria. Human evaluators also strongly favored our method for preserving object structure and generating physically plausible motion that reflects material properties.

MethodForce-direction following (%) ↑Physical plausibility (%) ↑Subject consistency (%) ↑
Ours78.576.480.5
AniClipart0.00.51.0
PhysCtrl16.414.412.8
Force Prompting4.17.74.6
Vector Prism1.01.01.0

Q3. Failure cases of the full pipeline

We provide three failure cases with the full pipeline active in Appendix A.4. On the snowman, a missed hat-head joint causes the hat to drift away and intersect the head under gravity. On the balloon dog, a force five times the default magnitude produces locally distorted PhysCtrl particle trajectories. Our projection preserves well-formed shapes and connectivity but cannot guarantee physically correct motion. On the balloon, rapid deformation of the soft string leads to inconsistent control-point motion during transfer, causing the path to fold over itself. Together, these cases show that projection can correct some geometric distortions, but cannot recover missingstructural constraints, guarantee physically correct trajectories, or prevent all artifacts introduced during vector-path transfer.

Thank you again for the constructive feedback. We hope the user study, multi-material results, and clarifications above address your concerns. We respectfully ask the reviewer to reconsider the rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear Area Chair,

We sincerely thank you and all three reviewers for the constructive discussion. We are pleased that all three reviewers now support acceptance (all rated 3: Weak acceptance), with Reviewer 4N5f raising their score from 2 to 3.

In summary, our rebuttal and revisions addressed all key concerns:

  • Reviewer 4N5f: We addressed baseline prompt sensitivity via neutral-prompt and comparable-motion-response evaluations, added a 30-participant user study and multi-material comparisons (Appendix A.3), and provided diverse force qualitative results.
  • Reviewer 98yp: We added quantitative guidance ablations (Table 3) and a cross-VLM sensitivity analysis with GPT-5 confirming the stability of our conclusions.
  • Reviewer jJLC: We added comparisons with Vector Prism (Figure 4, Tables 1 and 2), a detailed runtime breakdown (Appendix A.1), and stage-by-stage failure analyses (Appendix A.4).

We greatly appreciate your time and consideration in making the final decision.

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.