WorldFoundry
English中文

HY-WorldPlay

HY-WorldPlay

HY-WorldPlay from Tencent (Hunyuan): world and video; Verified.

Edit this pageGitHubMarkdown

HY-WorldPlay is Tencent's real-time interactive world video model: starting from an image, it follows pose/action input to roll the world forward — the interactive arm of the Hunyuan world stack. In this catalog the runner is verified for the official image-pose-video route with the official 8-GPU execution profile.

Official paper teaser: Overview of WorldCompass . 1) Starting from environmental prompts and action sequences, we generate shared prefix video clips. At the n n -th target clip, we perform clip-level rollouts to generate a set of candidate.
Official paper teaser: Overview of WorldCompass . 1) Starting from environmental prompts and action sequences, we generate shared prefix video clips. At the n n -th target clip, we perform clip-level rollouts to generate a set of candidate.

The recorded paper is arXiv 2602.09022.

Project page: 3d-models.hunyuan.tencent.com/world. Code: Tencent-Hunyuan/HY-WorldPlay.

Build your run

Pipeline
HunyuanWorldPlayPipelineworldfoundry.pipelines.hunyuan_world.pipeline_hunyuan_worldplay
TaskOfficial Image Pose Video official-image-pose-video
Environmentworldfoundry-unified-cu128
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  hy-worldplay \
  --pipeline.task-profile official-image-pose-video \
  --json

What it is for

Pick hy-worldplay when you need an action-conditioned world stepper: recorded actions and a frame in, a rollout out. Use it for interactive world inference, not a text-to-video prompt dump.

The recorded checkpoint set tells the assembly: tencent/HY-WorldPlay action weights over tencent/HunyuanVideo-1.5 video-model support assets, with Qwen2.5-VL-7B-Instruct, byt5-small, and FLUX.1-Redux-dev as support components. The in-tree pipeline follows the official demo contract. It is published by Tencent (Hunyuan).

Catalog tasks include world, video, and world model. Catalog aliases include hunyuan-worldplay and hunyuanworldplay. Documented outputs include video.

Recorded evaluation entry points include WorldModelBench (Interactive world-rollout quality), CameraBench (Pose/camera-following accuracy).

Runtime contract

Optional image, prompt, interactions, num_frames, and fps; artifacts hy-worldplay.mp4.

CLI task profiles: official-image-pose-video. WorldFoundry binds HunyuanWorldPlayPipeline.

Family status: Verified.

Recorded official-image-pose-video defaults: fps=24, height=480, width=832, seed=1.

The official route uses 8 torchrun ranks — plan for a multi-GPU host and stage all five recorded checkpoint sets. Full benchmark scoring and longer interaction suites remain separate work.

Running in WorldFoundry

Launch worldfoundry-eval run hy-worldplay --pipeline.task-profile official-image-pose-video to step the world. The command builder above switches --pipeline.task-profile among recorded tasks and still emits worldfoundry-eval run.

It runs in the unified environment worldfoundry-unified-cu128 (Python 3.11, CUDA 12.8).

Recorded weight repositories include tencent/HY-WorldPlay, tencent/HunyuanVideo-1.5, and Qwen/Qwen2.5-VL-7B-Instruct among others. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh hy-worldplay. Add --download to fetch missing public checkpoints.

Status: Verified.