PaperlayerView exact revision on arXiv

Source-linked AI summary

AnimateCanvas: Learning Implicit Motion Planning from Composable Kinematic Cues

Zeyu Ling, Di Kang, Qing Shuai, Yuxin Wen, Jing Li, Zhanke Wang, Heng Li, Chunchao Guo, Changqing Zou, Linchao Bao

arXiv:2609.10457v3cs.MM

TL;DR

Animation workflows need precise control over kinematic values that text-to-motion models do not specify, such as paths and joint orientations at chosen frames. AnimateCanvas learns full-body motion from heterogeneous cues on a shared canvas, and reports strong performance across temporal completion, spatial control, sequential generation, editing, and repair while retaining text-to-motion capability.

  • Problem

    Text-to-motion models offer semantic expressiveness but limited numerical control over paths and joint orientations at designated frames, despite animation workflows requiring both a strong motion prior and precise control.

    Source: §1 · p. 2

  • Method

    A shared flow-matching model uses a canvas for heterogeneous kinematic cues, cue imputation to preserve specified values, and compositional cue sampling to train across cue layouts.

    Source: §1 · p. 2

  • Results

    The model leads BABEL retrieval and transition-window FID, all MotionFix metrics, seven of eight automatic-repair metrics, and both PerMo tracks in target-attribute and paired-motion quality.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Composable kinematic cues provide a shared interface for controllable generation, sequential generation, language-guided editing, and motion repair while preserving text-to-motion capability.

    Source: §5 · p. 9

    The model covers 22-joint body motion without articulated hands or facial expression, and decoded non-root position control remains approximate without post-hoc inverse kinematics.

    Source: G LIMITATIONS · p. 49

Abstract

from arXiv · show

Professional character animation requires both natural motion and precise, versatile control. For example, creators often define the timing of a specified action, control the motion range of the character's arm swing, or specify the route the character walks through--effectively placing various kinematic cues on a motion canvas. This motivates us to propose AnimateCanvas, a model that supports cue-conditioned implicit motion planning to faithfully and coherently connect all cues, dense or sparse, full or partial, into one full-body motion sequence. Specifically, AnimateCanvas represents heterogeneous kinematic cues on a shared motion canvas, where position and rotation values are specified across body joints and time. A shared flow-matching model generates motion conditioned on this canvas, with optional language and input motion; cue imputation keeps the specified canvas values fixed in both training and sampling. To learn coherent completion across different cue sets, we train with a compositional cue sampler that varies the timing of cue application, the positions or rotations specified, and how they are combined. Together, these designs enable a single generator to integrate heterogeneous kinematic cues into coherent full-body actions, giving creators fine-grained control over selected frames, joints, and position or rotation channels. We evaluate this planning ability on temporal, root, and body-part cues--alone and combined--as well as language-guided editing, and naturally extend it to sequential generation and motion repair. AnimateCanvas achieves state-of-the-art results in temporal completion, spatial control, sequential generation, language-guided editing, and motion repair, while retaining strong text-to-motion capability.

1 INTRODUCTION

Animation workflows need precise kinematic control beyond text descriptions, while actions must connect cues across frames and body parts. AnimateCanvas uses a shared cue canvas, compositional sampling, and cue-preserving flow matching; evaluations report strong results across control and related motion tasks.

1 INTRODUCTION

  • Current text-to-motion models offer semantic expressiveness but limited numerical control, while animation workflows require specified paths and joint orientations at chosen frames.The motivating challenge is combining a strong motion prior with precise control over supplied kinematic values.Source: §1 · p. 2
  • Temporal cues anchor poses or retained segments at selected frames, while spatial cues prescribe paths, facing, joint positions, or local rotations, often in combination.Cues can apply at isolated instants or extended intervals and may preserve portions of an existing performance.Source: §1 · p. 2
  • The shared canvas records specified values and their frame, joint, and channel layout; a generation mask distinguishes specified from unspecified entries.Language and optional input motion provide separate contextual inputs, while compositional cue sampling varies frames, joints, and position or rotation channels.Source: §1 · p. 2
  • A shared flow-matching model uses cue imputation to fix specified entries and boundary and rollout objectives to learn transitions between cue-prescribed states.The model combines heterogeneous temporal and spatial cues through compositional sampling.Source: §1 · p. 2
  • AnimateCanvas leads BABEL retrieval and transition-window FID, all MotionFix metrics, seven of eight automatic-repair metrics, and both PerMo tracks in target-attribute and paired-motion quality.The same model also retains strong text-to-motion capability and supports sequential generation, editing, and repair.Source: §1 · pp. 2–3

2 RELATED WORK

Prior motion-control methods often use task-specific interfaces, while unified approaches and conditional-generation tools offer different forms of shared control. These lines of work motivate a shared completion behavior across heterogeneous cue layouts.

2 RELATED WORK

  • Task-specific generators address temporal completion, spatial targets, editing, style transfer, repair, and sequential or streaming generation.Their task-specific interfaces motivate a shared completion behavior across heterogeneous cue layouts.Source: §2 · p. 3
  • Unified systems combine different motion tasks through approaches including motion priors, joint-level imputation, source–condition–target mappings, and motion-language models.ANIMATECANVAS combines direct root-translation, rotation, and position cues with motion context across structured, irregular, and mixed layouts.Source: §2 · p. 3
  • Conditional-generation methods include known-value imputation, masked prediction, flow-path rewriting, kinematics-aware projection, and stochastic inference-time control.ANIMATECANVAS couples cue-preserving flow with learned geometric and temporal compatibility under mixed cue sets.Source: §2 · p. 3

3 METHOD

ANIMATECANVAS learns full-body trajectories from partial kinematic states using a shared canvas, cue-preserving flow matching, and geometry and transition supervision. Its training varies cue layouts and compositions so one model can complete motions across different specified frames, joints, and channels.

3 METHOD

  • A shared flow-matching model infers a full-body trajectory from the prescribed partial states using a motion canvas, cue sampling, cue-preserving flow, and geometry and transition supervision.Language and input motion can provide separate context for generation.Source: §3 · p. 4

3.1 MOTION CANVAS AND CUE COMPOSITION

  • The canvas stores cue values and a generation mask, with specified entries distinguished from entries to generate; language and input motion are separate context inputs.The motion representation includes root translation, 6D rotations of 22 joints, and pelvis-relative positions of 21 non-root joints.Source: §3.1 · p. 4
  • Compositional cue sampling varies selected frames and kinematic quantities, combining selections to train completion across structured and irregular cue layouts.Selected quantities include translation axes, joint rotations, and joint-position axes; correlated maps expose irregular layouts.Source: §3.1 · pp. 4–5

3.2 FLOW MATCHING FOR COMPOSITIONAL MOTION CONTROL

  • Cue imputation projects the initial state onto the specified canvas values, and the shared flow model learns target flow on unspecified entries while assigning zero flow velocity to specified entries.The conditional velocity model receives motion tokens containing state, source context, and mask, alongside language tokens; each Euler step restores the cues.Source: §3.2 · pp. 5–6

3.3 TRAINING OBJECTIVES

  • Geometry supervision compares forward-kinematic decoded joint positions and displacements with reference motion, while FK consistency couples position cues to rotations and translation.The time weighting emphasizes estimates made from less noisy flow states.Source: §3.3 · p. 6
  • Boundary losses target excess motion error near specified-to-unspecified transitions, while rollout supervision trains sampled clips to approach and depart isolated root-translation cues coherently.The rollout penalties supervise neighboring positions and second differences while allowing genuine turns and speed changes.Source: §3.3 · p. 7

4 EXPERIMENTS

The experiments assess temporal and spatial cue control, mixed cues, text-to-motion retention, editing, repair, artist judgments, and design ablations. Across these settings, ANIMATECANVAS combines cue adherence and motion-quality results, with some task-specific exceptions and trade-offs.

4.2 TEMPORAL CONTROL AND SEQUENTIAL GENERATION

  • ANIMATECANVAS has zero decoded cue error and the lowest FID in displayed text-conditioned temporal settings, while leading BABEL retrieval and transition-window FID.For sequential generation, it also leads reference-closest peak jerk, although its AUJ is 3.1× the lowest baseline.Source: §4.2 · p. 8

4.3 KINEMATIC CONTROL ACROSS SPATIAL SCOPES

  • Root-path ADE is 0 mm for dense paths and 30 mm for sparse waypoints; dense upper-body rotation error is zero and lower-body position error is 12.2 mm.Sparse-wrist XYZ error exceeds MotionLab, MaskControl, and ProjFlow.Source: §4.3 · p. 8

4.4 MIXED-CUE CONTROL

  • For mixed cues, ANIMATECANVAS improves FID, R@1–3, MM-Dist, and position adherence over CondMDI; full-composition training also outperforms singles-and-pairs training on held-out compositions.Without rotations, it leads FID/R@1–3, while ProjFlow exactly preserves positions.Source: §4.4 · p. 8
  • Optional model-independent IK raises P1–P4 all-position-cue success within 50 mm from 47.7% to 99.2%, adding 386 ms per motion with some quality trade-off.Main-table results exclude IK.Source: §4.4 · p. 9

4.5 TEXT-TO-MOTION GENERATION WITHOUT KINEMATIC CUES

  • Without kinematic cues, ANIMATECANVAS improves HumanML3D FID over HY-Motion-1.0-Lite from 0.0157 to 0.0108 and improves MM-Dist, with retrieval changing by at most 0.45 percentage points.The released MotionStreamer evaluator corroborates text-to-motion retention.Source: §4.5 · p. 9

4.6 INPUT-MOTION EDITING AND REPAIR

  • On MotionFix, ANIMATECANVAS leads reported metrics, raising M2M R@3 from 83.33% for the unedited source to 99.08%; it also leads PerMo attribute accuracy and paired-motion retrieval on both tracks.Automatic repair outperforms detector-driven StableMotion on seven of eight metrics, except acceleration.Source: §4.6 · p. 9

4.7 USER STUDY

  • Three technical artists gave positive net preference across every reported task and dimension for motion quality, physical artifacts, and repair effectiveness.They rated 1,218 anonymously ordered output pairs across four task types.Source: §4.7 · p. 9

4.8 ABLATION STUDY

  • The full model improves all six metrics over Basic completion; joint training-and-sampling imputation reduces Path ADE from 122 to 30 mm, and compositional sampling cuts joint-axis error from 81.1 to 40.9 mm.Cumulative objectives also reduce XZ/XYZ waypoint second differences from 13.34/13.46 to 1.62/1.96 mm and improve T2M FID from 0.0127 to 0.0108.Source: §4.8 · p. 9

5 CONCLUSION

The conclusion presents ANIMATECANVAS as a unified system for controllable motion generation and related tasks. The appendix provides implementation and evaluation details, while code and selected data products are slated for release.

5 CONCLUSION

  • ANIMATECANVAS uses a shared canvas, compositional cue sampling, and cue-preserving flow matching to infer full-body motion from heterogeneous kinematic cues.The paper presents the system as extending a pretrained text-to-motion model to controllable generation, sequential generation, language-guided editing, and repair while preserving text-to-motion capability.Source: §5 · p. 9
  • The authors state they will release training and inference code, model weights, evaluator code and weights, and processed test data, but not the non-redistributable training corpus.The manuscript was polished and engineering-oriented programming tasks assisted with generative AI tools; authors say they reviewed and verified the outputs.Source: REPRODUCIBILITY STATEMENT · p. 10, AI USE STATEMENT · p. 10

APPENDIX

  • The appendix separates implementation details, shared evaluation definitions, task protocols and results, design analyses, and artist assessments and limitations.Task protocols, results, and visual examples are grouped together in Sections C and D.Source: CONTENTS AND READING GUIDE · p. 13

A METHOD AND IMPLEMENTATION

The appendix details a shared motion representation and network, varied cue sampling, training objectives, optimization, and automatic repair. Its objectives address component balance and transitions around cues, while repair localizes anomalous regions before masked regeneration.

A.1 MOTION REPRESENTATION AND CONDITIONING

  • The 198-dimensional state retains root translation, rotations, and pelvis-relative positions; the model conditions on motion context and language and uses an 18-layer MMDiT to predict per-frame flow velocity.Motion context supplies editing-input values on unspecified entries and is zero elsewhere; motion attention spans at most 60 frames in either direction.Source: §A.1 · p. 14, §A.2 · p. 14

A.3 COMPOSITIONAL CUE SAMPLING DISTRIBUTION

  • Cue sampling varies selected frames, quantities, joints, axes, and combinations, including irregular layouts; cues sharing a canvas entry must prescribe the same value.Translation and position axes are independently selectable, while rotations use complete six-channel groups.Source: §A.3 · pp. 14–15

A.4 TRAINING OBJECTIVES

  • Training objectives balance active motion components and supervise decoded geometry, boundary behavior, root paths, and arrivals or departures around cues.Component balancing gives translation, root rotation, body rotation, and joint position equal weight; boundary terms compare join errors with generated-region baselines.Source: §A.4 · p. 15, §A.4.1 · p. 15, §A.4.2 · pp. 15–16

A.4.3 ROLLOUT SUPERVISION

  • Rollout supervision trains completed motion around isolated root cues, penalizing excessive path curvature and supervising neighboring positions and displacement changes.The rollout uses 50 imputed Euler updates, with gradients through the final four steps; excess is measured against the larger reference or generated-background level.Source: §A.4.3 · pp. 16–17

A.5 OPTIMIZATION AND SAMPLING CONFIGURATION

  • The model is initialized from HY-Motion-1.0-Lite and jointly trained across generation and editing tasks on approximately 700 hours of data for 2,000 epochs.Training uses 64 GPUs for approximately 14 days; inference uses classifier-free guidance 2.5 and 50 Euler steps with imputation.Source: §A.5 · p. 17

A.6 ANOMALY LOCALIZATION FOR AUTOMATIC REPAIR

  • Automatic repair compares corrupted motion with a self-denoised proposal to localize anomalies, expands repair regions across neighboring joints and frames, then regenerates masked entries.Unflagged input values are copied exactly, and reliable entries become fixed cues during regeneration.Source: §A.6 · pp. 17–18

B EVALUATION PROTOCOLS AND METRICS

  • The appendix organizes shared representations, evaluation definitions, task protocols and results, and design and artist analyses.Task protocols, results, and visual examples appear together in Appendices C–D.Source: B EVALUATION PROTOCOLS AND METRICS · p. 18

B.1 DATASETS AND COMMON REPRESENTATION

The evaluation uses HumanML3D for text-to-motion and kinematic control, and other named datasets for sequential generation, editing, and repair. A shared joint-position evaluator supports comparisons across representations, while the conversion check identifies a limitation of external-evaluator scores.

B.1 DATASETS AND COMMON REPRESENTATION

  • HumanML3D tests text-to-motion and temporal, root, body-part, and mixed control; BABEL, MotionFix, PerMo, and BrokenAMASS test sequential generation, editing, style/content editing, and repair, respectively.Outputs generally use neutral SMPL-22 at 30 fps; positions and translations are measured in millimeters and angles in degrees unless noted.Source: §B.1 · p. 18

B.2 JOINT-POSITION EVALUATOR

  • A separately trained, frozen joint-position evaluator maps generated and reference motions into a common space, avoiding inverse recovery of rotations from positions.All methods share evaluator weights, and benchmark test scores are not used for checkpoint selection.Source: §B.2 · p. 18
  • Table 5 covers prefixes, endpoints, boundary segments, and nonuniform keyframes, with and without text; ANIMATECANVAS preserves supplied poses in every setting and has lowest FID in text-conditioned rows.Figure 4 marks supplied frames in orange and generated motion in blue for contiguous, irregular, and nonuniform-keyframe completion layouts.Source: Table 5 · p. 19, Figure 4 · pp. 20–21
  • Embedding scores assess distributional fit and text/motion matching, while decoded geometric metrics assess cue execution and motion properties.The artist study supplies complementary perceptual evidence on five tracks.Source: §B.2 · p. 20
  • In the external evaluator’s 4,042-clip conversion check, converted ground truth has FID 67.4322 and R@3 0.9219, versus 0.0000 and 0.9442 for native ground truth.The conversion sensitivity motivates common-position evaluation but does not identify the cause of every cross-method ranking difference.Source: §B.2 · p. 20

B.3 EMBEDDING-BASED METRICS

This section defines embedding, cue-adherence, and physical diagnostics, then reports temporal completion and root-route results. The diagnostics distinguish motion quality, cue execution, and route accuracy.

B.3 EMBEDDING-BASED METRICS

  • Embedding metrics distinguish distribution fit from paired-match retrieval: FID compares embedding distributions, while R@k measures matches ranked among grouped candidates.Protocols differ across tasks, including 32-candidate T2M groups and multi-positive BABEL action groups.Source: §B.3 · p. 21

C.2 ROOT TRAJECTORIES AND WAYPOINTS

  • ANIMATECANVAS achieves zero dense-path ADE and failures at both thresholds, while sparse-waypoint ADE is 30 mm with zero full-path failures.The route score includes unsupplied frames between waypoints; ANIMATECANVAS also has the lowest FID and path ADE across XZ/XYZ routes.Source: Table 6 · p. 22, Figure 5 · p. 23, §C.2 · p. 29

B.4 CUE-ADHERENCE METRICS

  • Cue metrics score position or rotation error on supplied frames, joints, and axes; sparse root-path metrics also evaluate the inferred route between waypoints.Zero error for directly enforced constraints means satisfaction to numerical precision.Source: §B.4 · pp. 22–24

B.5 PHYSICAL DIAGNOSTICS

  • Foot, Jitter, and AUJ are proxies: high values alone do not establish severe visible skating or jitter, and reduced motion can also lower them.The complementary user study covers five tracks, not every diagnostic setting.Source: §B.5 · p. 24

C.1 TEMPORAL CONTROL

  • ANIMATECANVAS has zero cue error and the best FID/retrieval for prediction and in-betweening, with mixed results on nonuniform-keyframe embedding metrics.With text, CondMDI leads R@3 by 0.0003; without text, it leads embedding metrics, while ANIMATECANVAS preserves frames exactly.Source: §C.1 · p. 26

C.3 BODY-PART CONTROL

Body-part control tests sparse and dense position and rotation targets across body regions, joints, and axes. Mixed-cue protocols test whether one model handles position cues with and without rotation cues.

C.3 BODY-PART CONTROL

  • Body-part tests compare supplied-coordinate errors for sparse and dense positions and local rotations, with shared root and ancestor cues in upper/lower-body rotation comparisons.Isolated single-joint and bilateral rotation settings are reported for ANIMATECANVAS only because supplying reference-ancestor cues would change the input.Source: §C.3 · pp. 29–30
  • ANIMATECANVAS leads FID/R@3 in all 60 baseline-comparison body-part settings, and dense cues reduce its position error for all 14 XZ/XYZ targets.ProjFlow has lower position error in its reported settings, and FID gains do not uniformly improve Foot/Jitter.Source: §C.3 · pp. 30–31

C.4 MIXED-CUE COMPOSITION

  • Mixed-cue protocols vary position density and dispersed versus clustered layouts, comparing matched inputs with and without local-rotation cues.The same ANIMATECANVAS checkpoint is used in both groups; only the supplied cues change.Source: §C.4 · pp. 31–32
  • With rotation cues, ANIMATECANVAS improves quality and position adherence over CondMDI; without them, it leads FID/R@1–3 across P1–P4.ProjFlow exactly preserves all four position subsets in the no-rotation group; no cross-group ranking is inferred.Source: Table 10 · p. 30, Table 11 · p. 31, §C.4 · p. 32

C.5 ADDITIONAL BODY-PART EMBEDDING METRICS

The section adds body-part embedding metrics and evaluates the shared model in text-to-motion, sequential generation, and editing settings. Its sequential results cover action retrieval and transition quality, while editing protocols use paired source–target evaluations.

C.5 ADDITIONAL BODY-PART EMBEDDING METRICS

  • Table 12 supplements body-part comparisons with R@1, R@2, MM-Dist, and diversity alongside the primary tables’ FID, R@3, cue error, Foot, and Jitter.It covers body-region rotations and XZ/XYZ positions in sparse and dense layouts.Source: §C.5 · p. 32, Table 12 · p. 33

D.1 TEXT-TO-MOTION

  • Under shared joint-position evaluation, ANIMATECANVAS leads FID/MM-Dist over its initialization and published systems while retaining reference-like diversity.Against HY-Motion-1.0-Lite, retrieval changes by at most 0.0045 and MM-Dist improves from 27.3560 to 27.2810.Source: §D.1 · p. 33, Table 13 · p. 36

D.2 SEQUENTIAL GENERATION

  • ANIMATECANVAS leads BABEL subsequence retrieval/MM-Dist and transition-window FID, with peak jerk closest to the reference; AUJ remains higher than the best baseline.The table covers sequential generation on 1,295 BABEL episodes.Source: §D.2 · p. 36, Table 14 · p. 37

D.3 MOTION EDITING

  • MotionFix editing is evaluated on 1,013 input–instruction–target triplets, with models given only the input and instruction and targets reserved for evaluation.Paired retrieval and distance compare outputs with targets; the unchanged input is also evaluated against the same targets and candidate groups.Source: §D.3 · p. 37

D.3.1 ATTRIBUTE AND MOTION DIAGNOSTICS

  • PerMo style and action accuracy use five-way cosine retrieval against additive motion-embedding prototypes, while paired M2M retrieval separately measures complete-target agreement.PerMo physical diagnostics use second differences and report clip averages in millimeters without frame-rate normalization.Source: §D.3.1 · pp. 37–38

D.3.2 RESULTS AND QUALITATIVE EXAMPLES

ANIMATECANVAS edits motion in response to language and evaluates how cue sampling and transition objectives affect editing and repair. Results show strong target agreement, while the repair comparison distinguishes artifact reduction from exact-reference recovery.

D.3.2 RESULTS AND QUALITATIVE EXAMPLES

  • ANIMATECANVAS improves MotionFix source R@1 from 61.45% to 94.13%, M2M-Dist from 0.7260 to 0.2939, and FID from 0.0236 to 0.0056.The unedited source already exceeds editing baselines in R@3 (83.33%), so target agreement alone does not measure instruction application or attribute preservation.Source: §D.3.2 · p. 38
  • On MotionLab’s native evaluator, ANIMATECANVAS leads the reported metrics and raises source R@3 from 75.02% to 95.95%.This evaluator uses all 1,013 MotionFix triplets and corroborates target agreement.Source: §D.3.2 · p. 38
  • PerMo style edits retain action with S-Acc/A-Acc of 40.79%/92.54%, while content edits attain 48.21%/96.42%, versus their respective ground-truth values.The examples vary pose depth, limb orientation, timing, and starting side.Source: §D.3.2 · p. 39

D.3.3 EDITING WITH ADDITIONAL KINEMATIC CUES

  • With input motion, instruction, and seed fixed, examples add root paths, pose keyframes, or wrist rotations to language-guided edits.The figure contrasts editing alone with editing plus cues; these examples are outside the aggregate Table 15 evaluation.Source: §D.3.3 · p. 39, Figure 10 · p. 40

D.4 MOTION REPAIR

  • BrokenAMASS evaluates repair on 299 valid 100-frame clean–corrupted pairs, testing localization and regeneration separately and together.Known-region repair uses reference discrepancies only to identify anomaly locations; regeneration receives the corrupted clip and mask.Source: §D.4 · p. 40
  • Automatic repair beats detector-driven StableMotion on seven of eight metrics, except acceleration (3160 versus 3040 mm/s2).Against corrupted input, automatic repair reduces penetration, jitter, and acceleration but increases skating, MPJPE, and root error.Source: §D.4 · p. 42, Table 16 · p. 43

E.1 CONDITION INJECTION AND CUE SAMPLING

  • The named-task sampler uses a finite catalogue of temporal, root, body-part, and predefined mixed-cue templates rather than arbitrary template unions.Its templates draw cue values from training reference motions and leave unspecified entries masked.Source: §E.1 · p. 43
  • Full-model additions improve all six metrics over Basic completion; T2M FID changes from 0.0201 to 0.0108 and mixed joint-axis error from 102.6 to 40.9 mm.Relative to Full-state flow, imputation improves path adherence with similar sparse-waypoint FID.Source: §E.1 · p. 43

E.2 BOUNDARY AND ROLLOUT OBJECTIVES

Boundary and rollout objectives target waypoint transitions, and tests examine held-out cue compositions, optional inverse-kinematics refinement, and inference cost. The results show improved cue-local measures and control across several settings, alongside trade-offs in some whole-motion metrics.

E.2 BOUNDARY AND ROLLOUT OBJECTIVES

  • Cue-neighborhood losses supervise positions and one-sided displacement errors beside specified root cues, while the objective progression adds root-boundary, rollout, solver-endpoint, and background-relative excess terms.The table’s cumulative additions do not isolate each loss individually.Source: §E.2 · p. 44
  • The full objective gives the lowest waypoint second differences and Jitter in the cumulative progression, alongside improved T2M FID but worse diagnostic FID and retrieval.Rollout and solver-endpoint stages yield the largest further WP-∆2/C/B reductions.Source: §E.2 · p. 44, Table 17 · p. 45
  • Background-relative excess reduces XZ/XYZ WP-∆2 from 2.381/2.903 to 2.166/2.683 mm and C/B from 1.152/1.135 to 1.028/1.026.Retrieval changes marginally, while verification FID and XZ Jitter rise slightly.Source: §E.2 · p. 45, Table 17 · p. 45

E.3 GENERALIZATION TO HELD-OUT CUE COMBINATIONS

  • Full-composition training improves FID, R@3, and non-root accuracy in every protocol; withholding it changes Overall FID from 0.0092 to 0.0115 and R@3 from 0.9029 to 0.8898.The held-out test uses existing P1–P4 inputs; restricted-model performance supports generalization but not equivalence between seen and unseen compositions.Source: §E.3 · p. 46, Table 18 · p. 46

E.4 IK REFINEMENT OF POSITION CUES

  • IK keeps supplied root translations and local rotations fixed, and does not receive an unobserved reference pose, future joint target, or clean full-body motion.The refinement therefore adjusts generated motion under those equalities rather than relaxing them to reach position targets.Source: §E.4 · p. 46
  • Optional IK reaches 100% Success@50 mm for wrists and lower body and 99.2% for mixed cues, while reducing mean cue error, P95, and Max in every setting.FID rises by 0.0001–0.0006, and dynamics effects differ by task; position refinement trades off against some motion metrics.Source: §E.4 · p. 47, Table 19 · p. 47

E.5 INFERENCE EFFICIENCY

  • ANIMATECANVAS generates 360 frames in 2.47 seconds with 2.06 GB peak memory; its runtime is faster than listed 100-step KIMODO and 1,000-step diffusion/gradient-guided baselines, and close to MotionLab’s 2.06 seconds.The comparison uses native lengths and schedules, not parameter count alone.Source: §E.5 · p. 48, Table 20 · p. 48

F.1 STUDY DESIGN AND RATING PROTOCOL

  • Three professional technical artists compare outputs across five tracks, rating fluency, naturalness, pose quality, support, joint quality, penetration, and repair effectiveness on a five-point scale.One evaluator rated each case; the study includes sparse keyframes, constrained starts, multi-action transitions, loops, and repair.Source: §F.1 · p. 48

F.2 PAIRED RESULTS AND PERCEPTUAL DIAGNOSTICS

Artist comparisons favor ANIMATECANVAS on fluency in every track, but preferences vary across physical properties. The reported judgments are descriptive and do not estimate inter-rater agreement or cross-seed variation.

F.2 PAIRED RESULTS AND PERCEPTUAL DIAGNOSTICS

  • Fluency net preference is positive in every track (34.5–91.0 pp), while horizontal support ranges from 3.6 pp with 69.6% ties for constrained starts to 65.4 pp with 32.9% ties for repair.Repair effectiveness receives 194 better versus 3 worse judgments.Source: §F.2 · p. 49, Table 21 · p. 49
  • The 8,757 judgments share output pairs across dimensions and may share sources across cases; one of three artists rated each case, so inter-rater agreement and cross-seed variation are not estimated.Tasks and dimensions remain separate rather than pooled as independent observations.Source: §F.2 · p. 49

G LIMITATIONS

The paper’s limitations include the modeled body representation, position-control precision, cue compatibility, and the scope of interaction tasks.

G LIMITATIONS

  • ANIMATECANVAS models 22-joint motion without articulated hands or facial expression, and contradictory cues are outside scope.Root translations and local rotations are preserved exactly, but non-root position consistency is learned rather than hard-enforced; post-hoc IK reduces error but may increase Jitter.Source: G LIMITATIONS · p. 49

Related papers

Loading 2609.10457v3…