OMG Release Artifacts

Official release checkpoints and benchmark artifacts for OMG.

Paper Checkpoints

Model	Training step	File
OMG 50M	90,000	checkpoints/paper/50m/sstep=090000.ckpt
OMG 100M	100,000	checkpoints/paper/100m/sstep=100000.ckpt
OMG 300M	55,000	checkpoints/paper/300m/sstep=055000.ckpt
OMG 500M	50,000	checkpoints/paper/500m/sstep=050000.ckpt

Updated Checkpoints

The updated checkpoints were trained with diffusion loss only on the omni-modal LeRobot dataset.

Model	Training step	Status	File
OMG 100M	158,000	Recommended	checkpoints/updated/100m/sstep=158000.ckpt
OMG 300M	154,000	Experimental	checkpoints/updated/300m/sstep=154000.ckpt

The updated models improve the principal fixed-manifest text, audio-distribution, and HumanRef fidelity metrics over the paper 100M checkpoint. The 300M checkpoint remains experimental because simple-command rollout quality is not uniformly better than the updated 100M checkpoint.

The pretrained benchmark evaluator is stored at evaluator/step_004000.pt.

Benchmark Artifacts

• benchmark/samples/mixed_modalities_all_v1: fixed benchmark inputs.
• benchmark/results/mixed_modalities_all_v1: generated motions, tracker-executed motions, and metric outputs.
• SHA256SUMS: checksums for release checkpoints and the evaluator.

See the OMG benchmark documentation for evaluation commands and OMG-Data for training data.