DanceRAE

Pretrained reconstruction checkpoints for DanceRAE,

Code and inference instructions: PLAN-Lab/DanceRAE.

File	Frozen encoder
DanceRAE_dinov3.pt	DINOv3 ViT-L/16
DanceRAE_siglip2.pt	SigLIP2 ViT-L/16

Download

hf download onkarsus13/DanceRAE DanceRAE_siglip2.pt --local-dir /data/onkar2/DanceRAE
hf download onkarsus13/DanceRAE DanceRAE_dinov3.pt --local-dir /data/onkar2/DanceRAE

Each checkpoint requires its matching frozen encoder weights. These custom PyTorch checkpoints contain model state, EMA weights, architecture configuration, and training metadata; they are not Transformers from_pretrained model directories.

The current DanceRAE command-line inference loader uses the SigLIP2 checkpoint. Download its encoder separately:

hf download google/siglip2-large-patch16-256 \
  config.json model.safetensors preprocessor_config.json \
  --local-dir /data/onkar2/DanceRAE/pretrained/encoders/siglip2

The DINOv3 checkpoint is also distributed here. It requires the corresponding native DINOv3 encoder weights and runtime sources; the current SigLIP2 inference loader does not select this variant.