yangkaixing/ACE-Dance-Six-Tasks

ACE Dance Six Tasks

Checkpoints from the six ACE Dance runs started on 2026-11-12.

Task	Epoch	Optimizer updates	Inference weights	Status
reactive_l2f	500	10,500	ONLINE	Interim
reactive_f2l	500	10,500	ONLINE	Interim
responsive_l2f	2000	14,000	EMA	Final
responsive_f2l	2000	14,000	EMA	Final
music_to_leader	2000	50,000	ONLINE	Final
music_to_follower	2000	50,000	ONLINE	Final

l2f generates the follower conditioned on the leader. f2l generates the leader conditioned on the follower. The music_to_* tasks generate one dancer from music. Raw actor 0 is the follower and raw actor 1 is the leader.

Each native PyTorch snapshot includes the online and EMA model state, normalizer, and model configuration. Use the matching ACE runtime and task codecs. Reactive and music snapshots require the runtime's dataset module for the native normalizer. snapshot-manifest.json records the roles, sizes, and SHA-256 hashes.

The two Reactive checkpoints are milestones from ongoing runs. Held-out motion quality metrics are pending.