AnimajSAS/AIS_BI_LSTM_v0

Model Description

This is the AIS-BiLSTM model accompanying the paper:

https://arxiv.org/abs/2605.02742

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

The AIS-BiLSTM model takes as input a tensor of shape N×(D+1), where N is the number of frames in the sequence and D=596 is the pose dimension. The K input keyposes are placed at their target frame indices in the first D channels, with all other frames zero-filled. The last channel is a binary mask flagging the K keypose positions. The model predicts the full N×D pose sequence.

Model Details

Property	Value
Framework	PyTorch Lightning
Checkpoint format	.safetensors
Input	Masked sequence of shape N×(D+1)
Output	Completed sequence of shape N×D
Pose dimensionality	D = 596
Training sequence length	N=224 frames (fixed); variable length supported at inference
Frame rate	24 fps

Usage

See the GitHub repository

Training Data

Trained on the train split of the mib_rig_controllers_values dataset, comprising ~8.1 hours of professionally handcrafted animation at 24 fps.

Test Data

Evaluated on three test splits of the mib_rig_controllers_values dataset:

Split	Schedule	Description
test_held_out_algorithmic	DBA	Held-Out set with a deterministic keypose schedule from the domain-based algorithm
test_held_out_random	Random (r=0.9)	Held-Out set with a random keypose schedule
test_production	Block	5 min production data with ground-truth block keyposes

License

Apache 2.0