Text-guided dynamic 3D character generation has advanced rapidly, yet producing high-quality motion that faithfully reflects rich textual descriptions remains challenging. Existing methods tend to generate limited sub-actions or incoherent motion due to fixed-length temporal inputs and discrete frame-wise representations that fail to capture rich motion semantics. We address these limitations by representing motion with continuous differentiable B-spline curves, enabling more effective motion generation without modifying the capabilities of the underlying generative model. Specifically, our closed-form, Laplacian-regularized B-spline solver efficiently compresses variable-length motion sequences into compact representations with a fixed number of control points. Further, we introduce a normal-fusion strategy for input shape adherence along with correspondence-aware and local-rigidity losses for motion-restoration quality. To train our model, we collate BIMO, a new dataset containing diverse variable-length 3D motion sequences with rich, high-quality text annotations. Extensive evaluations show that our feed-forward framework BiMotion generates more expressive, higher-quality, and better prompt-aligned motions than existing state-of-the-art methods, while also achieving faster generation. Our project page is at: https://wangmiaowei.github.io/BiMotion.github.io/.
Figures & tables
Figure 2: Overview. BiMotion uses a B-spline representation for motion generation. During training (red arrow), vertex differences are converted into control points and encoded into motion latents. During inference (black arrow), the initial mesh and the text prompt generate motion latents that are decoded into control points and converted into the generated mesh sequence via B-spline reprojection.
Figure 3: B-spline VAE Pipeline. Given the initial shape (Pn,0,N0) and control points Pn,k , the Encoder compresses them into latent codes z0 and zk . The Decoder reconstructs the predicted control points P^n,k , which are then reprojected to point differences via the B-spline basis. Note, * indicates that Fk′ uses the FPS-sampled indices of F0′ , ⊕ denotes matrix addition, and ⊗ denotes matrix multiplication.
Figure 4: Qualitative Comparisons. Our method (BiMotion) results in superior motion quality and is more aligned with the user-provided text prompts. Artifacts for the baseline methods are highlighted in red. Please see the supplementary material for additional results.
Methods
Vbench
OC ↑
SC ↑
TF ↑
AQ ↑
DD↑
GVFDiffusion [ 116 ]
0.167
0.920
0.986
0.505
0.650
AnimateAnyMesh [ 101 ]
0.155
0.951
0.993
0.514
0.100
V2M4 [ 12 ]
0.175
0.876
0.986
0.478
0.750
BiMotion (Ours)
0.187
0.948
0.995
0.529
0.800
Methods
User Study
Efficiency
Table 1: Quantitative Comparisons. Our approach not only surpasses existing methods across multiple metrics ( Sec. 4.3 ), but also requires less time and peak GPU memory (PG) on an A100 GPU.
B-Spline
NF
Control-PE
LCorr
LRigid
Rec Error ↓
✗
✗
✗
✗
✗
3.237
✗
✓
✗
✓
✓
2.674
✓
✗
✓
✓
✓
1.328
✓
✓
✗
✓
✓
1.648
✓
✓
✓
✗
✓
1.303
✓
✓
✓
✓
✗
1.349
Table 2: VAE Quantitative Ablations. “Rec Error” indicates the mean L1 error ( ×10−2 ) averaged per trajectory and per instance.
Figure 5: B-spline Ablation. B-spline interpolation from predicted control points achieves lower L1 error over the entire sequence than linear interpolation on sampled raw differences.
Figure 6: Laplacian Regularizer L . B-spline ×2 interpolation comparison from a T=10 mesh sequence comparing Ridge [ 36 ] and our Laplacian regularizer L . L produces more natural results.
Figure 7: Normal Fusion (NF) Ablation. NF effectively separates nearby distinct parts. Without it, artifacts appear (see red).
Figure 8: Control-PE Ablation. Our control-PE (Bottom) captures fine motion ( e.g., the lion’s tail) and reduces artifacts (red circles) compared to conventional position encoding (Middle).
Figure 9: LRigid Ablation. Our local rigid loss reduces the red-circled artifacts during motion generation. Prompt: “The pistol tilts up to an angled pose, then returns to forward-facing.”
Figure 10: Robustness to Meshing. Ours produces consistent motion under mesh changes (Top to Bottom), whereas AnimateAnyMesh exhibits unstable motion and artifacts (red circles).
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
Figure A1 : Velocity Architecture. The network vθvel comprises S′ stacked layers with self- and cross-attention modules.
Figure A3 : Additional Qualitative Comparisons with AnimateAnyMesh. (a) AnimateAnyMesh [ 101 ] and (b) BiMotion (Ours).
Figure A4 : Puppeteer Qualitative Results. Results are rendered using the authors’ open-source renderer, whose texture appearance differs from ours and AnimateAnyMesh.
Figure A5 : Video Generation Model Comparison. Qualitative comparison of video outputs from (a) Adobe Firefly [ 1 ] , (b) HunyuanVideo-I2V [ 47 ] , (c) Wan-2.5 [ 92 ] , and (d) Kling 2.1 [ 46 ] . Note: these generation models convert transparent backgrounds to black by default.
Number of control points
4
8
16
32
L2 Error ( ×10−3 ) ↓
68.7
22.6
6.85
1.81
Appendix
Table A1 : Ablation of the number of control points.
Figure A6 : VAE Reconstruction Error vs. Sequence Length. Mean L1 reconstruction error for our VAE, computed per trajectory and averaged per instance, reported for each sequence-length bin.
Figure A7 : INR methods like DNF can introduce shape artifacts.
w/o N0
N0 Frequency-PE
N0 Concat.
Ours
1.328
1.331
1.294
1.078
Appendix
Table A2 : Additional normal-fusion ablations. Mean L1 reconstruction error ( ×10−2 ), averaged per trajectory and per instance.
Figure A8 : User Study Interface and Protocol. Left: Instructions page. Screenshot of the questionnaire introduction; personal or institutional information approved by the research ethics committee has been masked. Middle: Video comparison interface showing the conditions of both the first frame and the textual prompt, along with four videos generated by four randomly ordered methods. Right: Three evaluation questions used to collect user preferences, forming the basis of the subjective evaluation metrics.
Methods
TA ↑
MP ↑
ME ↑
GVFDiffusion [ 116 ]
2.343±0.407
2.300±0.380
2.443±0.387
AnimateAnyMesh [ 101 ]
2.314±0.394
2.686±0.464
2.443±0.357
V2M4 [ 12 ]
2.876±0.368
2.714±0.354
3.048±0.348
BiMotion (Ours)
4.095±0.334
4.062±0.328
4.048±0.339
Appendix
Table A3 : User study results. Each cell shows the mean ± standard deviation over all participants and examples.
Figure A9 : Auto-captioning pipeline. We propose an iterative captioning pipeline that incorporates both a caption inspector and a caption generator to produce accurate and consistent captions for dynamic assets.
Figure A10 : Frame-length cumulative distribution. Distribution of mesh sequence lengths (length ≤200 ) from DeformingThings4D [ 52 ] , ObjaverseV1 [ 23 ] , and ObjaverseXL [ 22 ] .
Figure A11 : VLM Captioning Ablation. We compare captions generated by three video–language models (VLMs). Using GPT-5 produces concise and motion-faithful descriptions closely aligned with the real 4D dynamics.
School of Artificial Intelligence and Data Science, University of Science and Technology of China · Suzhou Institute for Advanced Research, University of Science and Technology of China
Huazhong University of Science and Technology School of Computer Science and Technology Wuhan, Hubei, China · Zhejiang University School of Software Technology Ningbo, China