abidlabs/repro-olaf-world-spectral-assets
Olaf-World SeqΔ-REPA mechanism reproduction This self-contained scaled synthetic proxy tests the stated orientation mechanism, not full Olaf-World video training. It generates multiple observation contexts with rotated effect coordinates, compares a local-effect baseline against sequence-level alignment to an invariant temporal teacher, and measures held-out-context action transfer. python3 repro_olaf_spectral.py --seeds 12 --train 512 --test 2048 --steps 8 The official… See the full description on the dataset page: https://huggingface.co/datasets/abidlabs/repro-olaf-world-spectral-assets.
Olaf-World SeqΔ-REPA mechanism reproduction
This self-contained scaled synthetic proxy tests the stated orientation mechanism, not full Olaf-World video training. It generates multiple observation contexts with rotated effect coordinates, compares a local-effect baseline against sequence-level alignment to an invariant temporal teacher, and measures held-out-context action transfer.
python3 repro_olaf_spectral.py --seeds 12 --train 512 --test 2048 --steps 8The official repository is https://github.com/showlab/Olaf-World and its documented alignment variant uses a frozen V-JEPA 2 teacher. The full released pipeline needs pretrained checkpoints and GPU video-model infrastructure; this proxy is clearly labelled toy throughout.