WorldPlay

Use the existing CLI launchers after environment setup. Set the repository path, HunyuanVideo-1.5 model directory and input image in the launcher. Set action_ckpt in the selected JSON to the matching HY-WorldPlay weights.

Model	Launcher	Config	Example pose
Autoregressive	run_worldplay_ar.sh	worldplay_ar_i2v_480p.json	d-31
Bidirectional	run_worldplay_bi.sh	worldplay_bi_i2v_480p.json	s-31
Distilled	run_worldplay_distill.sh	worldplay_distill_i2v_480p.json	w-31
Autoregressive, four GPUs	run_worldplay_ar_sp4.sh	See launcher	d-31

Run from the repository root:

bash scripts/worldplay/run_worldplay_ar.sh

The launchers use seed 1 and save videos under save_results/HY-WorldPlay. For integration with Python, use the same config and request fields:

import json

from lightx2v.models.runners.runner_factory import build_runner
from lightx2v.utils.set_config import build_startup_config

with open("configs/worldplay/worldplay_ar_i2v_480p.json") as f:
    config = json.load(f)
config["model_path"] = "/path/to/HunyuanVideo-1.5"
config["action_ckpt"] = "/path/to/HY-WorldPlay/ar_model/diffusion_pytorch_model.safetensors"
runner = build_runner(build_startup_config(config))
request = runner.prepare_request({
    "seed": 1,
    "prompt": "A paved pathway leads towards a stone arch bridge.",
    "image_path": "/path/to/image.png",
    "pose": "d-31",
    "save_result_path": "/path/to/worldplay_ar.mp4",
    "return_result_tensor": False,
})
result = runner.run_request(request)

For the bidirectional and distilled models, change the config, action checkpoint and pose together using the table above. The bidirectional example accepts an empty negative_prompt; omit it for CFG-disabled models.