EgoRecover models and current continuation states

Code, protocols and conclusions: https://github.com/sxh-kk/EgoRecover Data and prepared caches: https://huggingface.co/datasets/sxhkk/EgoRecover-data

This release contains 158 files (25.625 GiB logical size): 24 formal G best checkpoints from T0–T3 and R0–R3, historical P/G comparisons, 41 small initial checkpoints needed by paired-control checks, the original E7 initializer including its optimizer-stored EMA, and six continuation states for R0 / expanded P seeds 62–64. Identical file hashes identify shared content; source relative paths are preserved.

The current inference combination is R0 seed62/63/64 best + exp/egorecover_g_pretraining/v1/prior/prior.pt (calibrated P), 1.096 GiB total. The raw D_s62 P is a different checkpoint and must not replace calibrated P for G reproduction. R0 Body MPJPE: 131.913±2.465 mm; matched E7 reference: 115.018 mm. These are a repeatedly used causal project benchmark, known floor and fixed startup shape, dev-selected best, Euler10 and one draw1062; they are not the original paper's independent-window protocol. See the GitHub results for full limitations.

release-index.json records each path, SHA256, size, purpose and download preset. python tools/restore_release.py --preset core restores the current inference weights plus normalization stats. The g-evaluation, p-evaluation, training and all presets restore their required artifacts. The GitHub index pins exact HF revisions after publication.

SMPL-X assets are not included. Obtain an authorized model separately. Source checkpoint identities, including original machine paths, remain unchanged to preserve hashes. The dedicated reproduction entry resolves the expanded P's legacy data pointer at runtime. Complete bitwise training reproduction on another machine is not claimed.

Original method: UniEgoMotion (ICCV 2025), Chaitanya Patel et al.; E7 adaptation: https://github.com/sxh-kk/UEM-update at 156ab79d5f692a6e8db3c3eb769abf1e4cb1fc08. Keep upstream notices and follow the terms for each source/asset; this card does not grant a new license over third-party assets.

Large prepared archives and checkpoints are stored as ordered, content-addressed parts of at most 64 MiB. Use the GitHub restore utility to assemble exact original bytes and verify both part and whole-file SHA256. Original large .pt/.zip paths are logical restoration paths; see parts in release-index.json for physical Hub objects.

Publication complete. All selected transport objects are committed; exact sizes and SHA256 are verified against remote metadata. A clean public-Hub g-smoke restoration and CPU inference passed. GitHub docs/release/artifacts.json pins the final revision.