# OMG-Lite 38D · native 50 Hz · best validation step 75500 This is the 38D checkpoint from the 50 Hz, text-only run, selected by validation loss at step 75500. It uses diffusion loss only (FK auxiliary loss weights are zero). It is not the original 125D model. ## Download and run Use Python 3.10 and a CUDA-capable PyTorch installation. The tested source revision is pinned below. From any working directory: ```bash git clone https://github.com/KunYing-Lee/OMG-Lite.git cd OMG-Lite git checkout f8d5be45979b0ddf4e4cdd49b9a512179200d94f python3.10 -m venv .venv source .venv/bin/activate python -m pip install -e '.[train]' 'huggingface-hub>=0.34,<1' python - <<'PY' from huggingface_hub import snapshot_download snapshot_download('THU-MARS/OMG', allow_patterns=['checkpoints/38D/*'], local_dir='hf-artifacts') PY python hf-artifacts/checkpoints/38D/infer.py \ --text 'walk forward' --output outputs/walk_38d.npz ``` The first inference downloads **google-t5/t5-base** tokenizer/config/weights. For an offline machine, download that model beforehand and pass `--t5 /absolute/path/to/t5-base`. The checkpoint includes the encoder state; strict checkpoint loading restores its weights. T5-3B is used by a separate benchmark evaluator and is **not** the generator's text encoder. The inference package does not require the training dataset or a tracker. `infer.py` resolves assets relative to its own location, verifies the checkpoint and supporting file SHA-256 values, strictly loads the state dictionary and validates the architecture contract. It invokes the repository's `CheckpointPlanner` without CUDA graph capture, with 50 DDIM steps, CFG 2.5, BF16 CUDA inference and seed 1234 by default. No smoothing, foot correction or trajectory alignment is applied. ## Input and output The model predicts 100 future frames (2 seconds) from ten native-50Hz history frames. The included `seed_stand.npz` supplies a standing startup history. To condition on your own motion: ```bash python hf-artifacts/checkpoints/38D/infer.py \ --history /absolute/path/to/history.npz \ --text 'squat down and stand up' --output outputs/squat_38d.npz ``` History NPZ: `qpos_36` shaped `[T,36]`, T >= 10, finite float values, scalar `fps=50`. The final ten frames are used. Layout: world root XYZ (metres), root quaternion WXYZ, 29 joint angles (radians) in the released G1 kinematics order. Do not pass 38D normalized features as qpos. Output NPZ: `qpos_36 [100,36]`, scalar `fps=50`, `caption`, `history_qpos_36 [10,36]`, `seed`. These are raw generated kinematic references, not a rendered video or tracker rollout. ## Configuration and statistics `config.yaml` is the saved model subtree with only three path substitutions: bundled stats, bundled kinematics, and portable T5-base model location. All saved architecture, normalization, diffusion and conditioning settings are preserved. Training dataloaders, callbacks and server paths are omitted. Optimizer/scheduler fields inside the model are historical constructor metadata, unused during inference; this config is not a recipe to resume the extended training schedule. - 38D = root translation 3 + root rotation 6D + 29 joints. Body positions are computed by FK. - Hidden dimension 512, 10 layers, 8 heads; sinusoidal history encoding and self/cross-attention QK normalization. - Exact training normalization: `g1_38d_50hz_hist10_stats.json`; history10, canonical frame9, horizon100, 50Hz. Do not substitute the 125D or original 30Hz stats. - Frozen T5-base in eval mode, maximum text length 50, cache size4096. To use `config.yaml` with the existing simulation CLI, set absolute asset paths through the environment: ```bash export OMG38_ARTIFACT_DIR="$(pwd)/hf-artifacts/checkpoints/38D" # Optional for offline use: # export OMG38_T5=/absolute/path/to/t5-base python -m omg.cli.pipeline.main \ --config "$OMG38_ARTIFACT_DIR/config.yaml" \ --checkpoint "$OMG38_ARTIFACT_DIR/sstep=075500.ckpt" \ --seed-motion "$OMG38_ARTIFACT_DIR/seed_stand.npz" \ --tracker /absolute/path/to/model_16200_dynamic_batch.onnx \ --text 'walk forward' --mode sync --frames 500 --output outputs/sync_38d ``` Simulation additionally requires the OMG-Lite tracking dependencies and the separately obtained HoloMotion v1.4.1 tracker. Follow the pinned repository's installation instructions. Raw generation above does not need these dependencies. ## Provenance and limits Training source: `ede4eb758b1b7a21e4c6227a6ff9641b1ba1d271`. Tested inference source: `f8d5be45979b0ddf4e4cdd49b9a512179200d94f`. Training was extended from three to five epochs at step41000 with a revised remaining learning-rate schedule; this checkpoint is at approximately 4.58 epochs. See `provenance.json` and `validation.json` for hashes and the actual inference check. On the fixed 1024-text, 20-seed comparison, this checkpoint improved text matching over the earlier 125D step49431 model but worsened FID and some raw-motion physical metrics. It is not an unconditional quality upgrade or evidence of physical-robot readiness.