OneWorld merged inference weights

One complete model.safetensors contains the generator EMA, RAE with final decoder EMA, and text-projector EMA, stored under their original parameter names. All values retain their original precision. There are no rank-local model shards.

Use the matching repaired inference code supporting oneworld-4step-v3, with the public Wan2.1-T2V-1.3B and Pi3X backbones. The compatible code is published in GitHub commit ebd6f31; follow its setup and inference instructions. The default HF main release is ready to use directly, without a repair or merge step. No input-alignment initializer or decoder shard-layout file is needed. Latent statistics are stored separately in latent_stats.pt.

Verify with python infer/verify_weights.py --model /path/to/this/package, then use the same path with python infer/infer.py --model ....