[CVPR2026 Highlight] BiMotion: B-spline Motion for Text-guided Dynamic 3D Character Generation

📑 0. Open-Source Plan

We plan to release all components of our project according to the following schedule:

• ✅ Paper release – before 25/02/2026
• ✅ Project page setup – before 25/02/2026
• ✅ Inference code & model weights – released
• ✅ BIMO dataset – released
• ✅ Training scripts – released

📝 1. Abstract

Text-guided dynamic 3D character generation has advanced rapidly, yet producing high-quality motion that faithfully reflects rich textual descriptions remains challenging. BiMotion addresses 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.

Key Contributions:

• B-spline Representation: A closed-form, Laplacian-regularized B-spline solver that efficiently compresses variable-length motion sequences into compact fixed-length control points.
• Enhanced Geometry Adherence: Introduces a normal-fusion strategy for input shape adherence along with correspondence-aware and local-rigidity losses.
• BIMO Dataset: A new dataset featuring diverse variable-length 3D motion sequences with rich, high-quality text annotations.

🛠️ 2. Environment Setup

Tested Environment: CUDA 12.4 / 12.8, Ubuntu 22.04, Python 3.10

2.1. Basic Installation

# Clone Repository 
git clone --recursive https://github.com/wangmiaowei/BiMotion.git
cd BiMotion

# Create Environment
conda env create -f environment.yml
conda activate BiMotion
pip install -r requirements.txt

2.2. Specialized Dependencies

Installation of pytorch3d and flash-attention can be sensitive to environment versions. We recommend:

# PyTorch3D
pip install --no-build-isolation "git+https://github.com/facebookresearch/pytorch3d.git@V0.7.8"

# Flash Attention (Recommended pre-built wheel)
pip install https://github.com/Dao-AILab/flash-attention/releases/download/v2.8.3/flash_attn-2.8.3+cu12torch2.5cxx11abiFALSE-cp310-cp310-linux_x86_64.whl

If you encounter GLIBC errors, build from source:

pip uninstall flash_attn -y
git clone https://github.com/Dao-AILab/flash-attention.git
cd flash-attention
pip install ninja
pip install --no-build-isolation .

2.3. Blender Integration

pip install bpy==3.6.0 --extra-index-url https://download.blender.org/pypi/

wget https://download.blender.org/release/Blender3.6/blender-3.6.0-linux-x64.tar.xz
tar -xf blender-3.6.0-linux-x64.tar.xz
rm blender-3.6.0-linux-x64.tar.xz

🚀 3. Evaluation & Inference

📂 3.1. Download Weights

Download the pre-trained checkpoints from Hugging Face:

WEIGHT_DIR="/path/to/your_model_weights"

hf download miaoweiwang/BiMotion \
  --repo-type dataset \
  --local-dir "$WEIGHT_DIR" \
  --include "model_weights/*.pt"

🏃 3.2. Running Inference

We provide two modes for motion generation depending on your input format:

Option A: 3D textured Object (.glb)

Generate animated results from a static GLB file.

python demo_evaluate_glb.py \
  --glb_name man \
  --prompt "He pivots, rotates his torso, sweeps one leg in a circular kick, then plants his foot and returns to guard." \
  --guidance_scale 3 \
  --export_format fbx \
  --seed 0 \
  --input_glbs_folder "visualize_experiments/input_glbs" \
  --save_path "visualize_experiments/results" \
  --vae_path "$WEIGHT_DIR/model_weights/vae_weight.pt" \
  --dit_path "$WEIGHT_DIR/model_weights/diffusion_weight.pt" \
  --azi 90 --ele 45 --generated_seq_length 20

Note: glb_name is the filename in your input folder (exclude the .glb extension).

Result: You will find man_azi90_ele45.mp4 in your results folder as follows.

Option B: Untextured raw Mesh (.obj / .ply)

Suitable for meshes without textures.

python demo_evaluate_mesh.py \
  --mesh_name "hand" \
  --prompt "The fingers bend toward the palm and then straighten out." \
  --guidance_scale 3 \
  --seed 0 \
  --input_meshes_folder "visualize_experiments/input_meshes" \
  --save_path "visualize_experiments/results" \
  --vae_path "$WEIGHT_DIR/model_weights/vae_weight.pt" \
  --dit_path "$WEIGHT_DIR/model_weights/diffusion_weight.pt" \
  --generated_seq_length 20

Note: The system automatically detects the mesh extension.

Result: You will find 20 sequential mesh files in the output path.

💡 3.3. Tips for Better Results

• 🎲 Try Multiple Seeds — Outputs vary; experiment with different seeds.
• 🎚️ Adjust Guidance Scale — Increase to 5.0 if motion is too subtle.
• ✍️ Refine Prompts — Rephrase or use keywords similar to BIMO annotations.
• 🧭 Check Mesh Orientation — Ensure Z-up and facing -Y.

📊 4. Model Training

4.1. Download the BIMO Dataset

# Set the root directory where the dataset will be stored
# (Ensure you have ~3.5 TB of free space)

DATASET_PATH="YOUR_DATASET_PATH"

python dataset/download_dataset.py $DATASET_PATH

• DATASET_PATH should be the root directory where the BIMO dataset will be saved.

• Each motion subdirectory retains its original dataset name:

  • DeformingThings4D (DT4D)
  • UID from ObjaverseV1 (paper)
  • SHA256 from ObjaverseXL (website)

Note: For the training steps below, we provide a SLURM script for single-node, multi-GPU training. For multi-node, multi-GPU setups, we recommend using 👉 idr_accelerate.

4.2. Train the B-spline VAE Model

1. Open the training script and modify the parameters that are already clearly marked:

scripts/script_vae_bspline.sh

2. Launch training (example: 1 node with 8 GPUs):

sbatch scripts/script_vae_bspline.sh

4.3. Compute Global Latent Statistics

We can compute the global latent statistics (mean and std) using 1 GPU. You may also run it via srun on a SLURM system if preferred.

#!/bin/bash

# Launch encoding of latent statistics
accelerate launch --num_processes 1 encode_latent_statistic.py \
  --batch_size 32 \
  --use_fp16 \
  --weight_decay 0.01 \
  # 🔴 TODO: set the checkpoint path of your trained VAE model
  --ckpt YOUR_EXPERIMENT_PATH/Bspline_motion_generation/VAE_Training/checkpoints/Your_Latest.pt \
  --config configs/vae_animate.yml \
  --start_idx 0 \
  --end_idx 38727 \
  --txt_file dataset/full_paths_V2.lst \
  # 🔴 TODO: set the root path to your dataset
  --data_dir YOUR_DATASET_PATH/bspline_motion_dataset/

Notes:

• Ensure the --ckpt points to the same checkpoint used in VAE training.
• Ensure the --data_dir matches the one used during VAE training.
• Batch size can be adjusted based on your GPU memory.
• This script saves motion and static latent mean/std in dataset/data_statistics/.

4.4. Train the Diffusion Model (DIT)

1. Open the training script and update the marked parameters:

scripts/script_dit_bspline.sh

2. Launch training (example: 1 node with 8 GPUs):

sbatch scripts/script_dit_bspline.sh

Notes: Keep the trained VAE checkpoints the same as step 4.3 when providing input to the diffusion model.

📜 Citation

If you find BiMotion useful in your research, please cite:

@article{wang2025bimotion,
  title={BiMotion: B-spline Motion for Text-guided Dynamic 3D Character Generation},
  author={Miaowei Wang and Qingxuan Yan and Zhi Cao and Yayuan Li and Oisin Mac Aodha and Jason J. Corso and Amir Vaxman},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year= {2026},
}