sxhkk/UniEgoMotion-e7

UniEgoMotion E7 — 88M, full-body / 全量关节

Original dense E7 training checkpoint, trained for 300 epochs (zero-based epoch 299), 84,000 optimizer steps. The training log reports 88.0M total parameters, including 87,450,099 trainable parameters.

原始 E7 全量关节训练权重。输出完整 243D v4_beta 表示，包含 22 个身体关节、全局运动、手部 PCA 与体型等信息。该权重不是 K12 稀疏模型或 400M 模型，也不是官方 Diffusion 权重。

Files

• last.ckpt: unmodified original PyTorch Lightning training checkpoint; includes model, EMA, optimizer, and scheduler states (1,417,277,447 bytes).
• e7_x0_global_w8_u84k.yaml: original E7 experiment configuration.
• checkpoint_metadata.json: verified checkpoint metadata and code reference.
• SHA256SUMS: SHA-256 checksum of last.ckpt.

Model and training

• Data: processed EE4D-Motion; DINOv2 image features and Aria trajectory conditioning.
• Motion window: 80 frames; full output shape [B, 80, 243].
• Objective: Flow Matching, clean-motion (x0) prediction.
• Global SE(3) feature range: [198:207], loss weight 8.
• Sampling: Euler, 10 steps.
• Original training: batch 256 per GPU, 2 GPUs, 300 epochs, learning rate 8.5e-5, EMA decay 0.992028.

EVAL.KEY_JOINTS_ONLY: True limits evaluation metrics to a 12-joint subset; it does not reduce the model's full-body output.

Download and use

hf download sxhkk/UniEgoMotion-e7 --local-dir exp/e7

Use the E7 codebase. The reviewed code revision is aebdbe25b93e37dd2f6d51bef4e4c00e312b5c5c. Install its dependencies and prepare the conditioning data and required SMPL-X assets as described there.

From that codebase, with a prepared conditions.pt:

python -m run.sample_e7 --checkpoint exp/e7/last.ckpt --conditioning conditions.pt --output output/e7.pt --repaint off --seed 62

conditions.pt is user-prepared input, not included here. Outputs are normalized motion representations; decoding requires the training normalization statistics and the codebase's motion decoder. Dataset files and SMPL-X model assets are not included.

Raw versus EMA weights

The checkpoint uses ema_state_format = original_model_with_ema_optimizer_v1:

• state_dict contains original training model weights, with the model. prefix.
• optimizer_states[0]["ema"] contains the ordered EMA tensors for trainable parameters.

Use the codebase's checkpoint loading / EMA helper to reproduce EMA-based evaluation. Loading only state_dict does not apply EMA. Preserve the complete checkpoint when resuming the original training run.

Verification and provenance

The original checkpoint was read on CPU to verify its epoch, optimizer step, representation, input/output projection shapes, and EMA state. This upload preserves the original checkpoint bytes. No new training or real-data evaluation was performed for this upload, and end-to-end inference against the latest E7 branch was not rerun.

Code: UEM-update E7, derived from UniEgoMotion, Chaitanya Patel et al., ICCV 2025. Refer to the source repositories for citation and applicable code, data, and model terms.