moge-UMO-ST-reaction-v5-2

Moge UMO ST: Reaction v5.2

Research checkpoints for the unified K273 motion model trained for text-to-motion, semantic motion editing, and source-conditioned human reaction generation.

Code and the complete experiment history are available at:

• Repository: https://github.com/CHDTevior/moge_UMO_ST
• Release branch: agent/reaction-v5-2-all-t-fine
• Release commit: db09c588e2f642e4809302d256361fac48079e33

Checkpoints

Both files are full PyTorch research checkpoints. They contain model weights, EMA weights, optimizer state, deterministic scheduler state, normalization statistics, and the resolved training configuration. Inference should use the ema state unless a specific diagnostic requires raw model weights.

Training schedule

0K-100K:    T2M / Edit / Reaction = 100 / 0 / 0
100K-300K:  T2M / Edit / Reaction = 30 / 35 / 35

Reaction v5.2 changes the v5.1 objective by activating the fine interaction and contact losses at every valid flow timestep (min_flow_t=0.0 and fine_min_flow_t=0.0). The backbone, data mixture, loss weights, optimizer, EMA, and inference protocol remain unchanged from the controlled v5.1 arm.

The 300K checkpoint has not been trained for Kimodo-style control or Ease conditioning. It should not be described as a control checkpoint.

Inference guidance

• T2M: EMA, ODE32; evaluate CFG values under one fixed protocol.
• Editing: source CFG 2.0; for this 300K model, edit CFG 2.0 is preferable to 3.0 on the inspected temporal-edit subset.
• Reaction: source CFG 2.0, text CFG 2.0, fixed source actor role.

The model expects the repository's cached LLM2Vec Llama-3 8B sentence and context-token features. Dataset files, text caches, evaluator weights, and training data are not included in this model repository.

Loading

import torch

checkpoint = torch.load(
    "checkpoints/reaction_v5_2_300k.pt",
    map_location="cpu",
    weights_only=False,
)
ema_state = checkpoint["ema"]
config = checkpoint["config"]

Use sample_hy273_multitask.py and the task-specific launch/evaluation scripts from the release branch to construct the model and conditioning tensors.

Main result boundary

At 300K, v5.2 improves Reaction geometry, relative heading, bearing, partner-facing accuracy, and contact-vector accuracy over the matched v5.1 300K arm. It does not solve precise contact-state or contact-transition timing. The release contains one training seed; reported bootstrap intervals cover evaluation examples, not training-seed variance.

See reports/HY273_REACTION_V5_2_ALL_T_FINE_300K_FINAL_REPORT_CN_20260809.md for the full Chinese report and metric definitions.