Adaptive Interpolation-Synthesis for Motion In-Betweening on Keyframe-Based Animation

Official implementation of the method described in the following paper, including training and evaluation scripts:

Raël et al., "Adaptive Interpolation-Synthesis for Motion In-Betweening on Keyframe-Based Animation", SIGGRAPH 2026 Conference Papers (2026)

• Installation
• Setup
• Training
• Evaluation
• Development
• Citation

This repo uses uv as the package manager. See the uv documentation for installation instructions.

Once installed, run:

uv sync --frozen --all-groups

Create a .env file at the repo root:

# Dataset & experiment directories
export MIB_POCOYO_DATASET_DIR=data/dataset
export MIB_POCOYO_EXPERIMENT_DIR=data/exp

# Optional: MLflow tracking server
export MLFLOW_TRACKING_URI=https://your-mlflow-server-uri

If the dataset directory is missing, it will be automatically downloaded from AnimajSAS/mib_rig_controllers_values on HuggingFace.

uv run python -m motion_inbetweening.scripts.train pocoyo --training-type best

Trained in ~2 hours on a RTX 4070.

uv run python -m motion_inbetweening.scripts.test path/to/checkpoint
# e.g. uv run python -m motion_inbetweening.scripts.test data/exp/<exp_name>/checkpoints/safetensors

path/to/checkpoint should point to the safetensors directory, generated at the end of training.

We provide a pretrained model on HuggingFace at AnimajSAS/AIS_BI_LSTM_v0:

uv run python -m motion_inbetweening.scripts.test AnimajSAS/AIS_BI_LSTM_v0

By default, all three test sets are evaluated. Use --test-set to run a specific one:

uv run python -m motion_inbetweening.scripts.test path/to/checkpoint --test-set held_out_algorithmic

Valid values for --test-set:

• held_out_algorithmic — in-house dataset with algorithmic block schedule
• held_out_random — in-house dataset with random masking
• production — production test set
• all (default) — runs all three test sets

Run linting and formatting checks before committing:

uv run pre-commit run

This project is licensed under the Apache 2.0 License.