This guide covers Tencent Hunyuan HY-WorldPlay, the HY-World 1.5 research project. It is separate from payment services and unrelated companies with similar names. Start with the official repository for code and checkpoints.
What WorldPlay generates
The project describes action-conditioned streaming video from an image or text prompt. Its goal is to preserve a scene as the viewpoint changes. That is different from receiving an editable polygon mesh, a verified physical simulation or a conventional game level. This guide does not demonstrate a model run.
Choose a repeatable scene before generating
Our fictional exercise is a gallery corridor with a yellow doorway, a coral chair and a tiled floor. Those landmarks make changes easier to identify than a scene filled with similar foliage. Write down their relationships before requesting movement: the chair sits near the doorway, and the courtyard is visible beyond it.
Use an owned reference image with readable architecture and enough depth to make the planned camera movement meaningful. Keep the original file and its crop with the prompt. The separate cover illustrates a fictional gallery; it is independent editorial artwork, not a WorldPlay result or evidence of recovered geometry.
Check the pipeline and hardware first
The checked README distinguishes HunyuanVideo-based and lighter WAN pipelines, with different quality tradeoffs. Its documented AR-distilled 125-frame setup lists memory of 72G at sp=1, 34G at sp=4 and 28G at sp=8. These are project configuration figures, not our measurements or a guarantee for your GPU.
Use the pinned setup documentation to check dependencies, model downloads and checkpoint paths. Record the selected pipeline before comparing a result. Do not combine a lightweight configuration’s requirements with another configuration’s quality or speed claim.
Plan pose segments with the local calculator
In the documented pose-string format, durations count latents, not seconds. One initial latent is included. Our original calculator checks that planning arithmetic and the documented frame divisibility rules. It does not install WorldPlay, download weights, run inference or measure camera motion.
Download the Python pose planner. It uses only the Python standard library and makes no network requests. After saving it, run:
python3 worldplay-pose-planner.py --pose 'w-15,s-15,right-1' --model ar --frames 125The supplied example contains 31 action latents plus the initial latent, giving 32 total and 125 frames under the checked convention. It plans forward movement, backward movement and a rightward turn. It does not prove an exact return to the original camera position.
Download the computed example report. This JSON is our planning report, not a provider camera-trajectory file. Do not pass it as the repository’s custom pose JSON input. Use the official trajectory generator when you need that separate format.
Understand what the calculator rejects
For the documented convention, frame count is four times total latents minus three. Autoregressive planning requires total latents divisible by four; bidirectional planning requires divisibility by sixteen. A 31-action-latent example satisfies both. A three-action-latent example satisfies the former but not the latter.
The calculator rejects an unknown action, zero duration, incompatible supplied frame count or failed divisibility check. It also imposes its own 4096-total-latent guard to keep this small planning utility bounded. That guard is not a provider capacity claim. Successful arithmetic does not validate every model setting, checkpoint or custom trajectory.
Review landmarks after moving away and back
Download the blank landmark review sheet. Record actual returned frame numbers or times after you have an output. The sheet has no invented observations or model score.
Compare the doorway’s structure, chair placement and floor pattern at the opening, approach and return. Note whether the chair changes shape, the doorway shifts or previously visible objects disappear. A visually plausible new frame can still contradict an earlier view of the same space.
Separate instruction following from appearance. First ask whether the requested movement occurred. Then ask whether the world stayed coherent. Finally inspect the file’s actual frame count, dimensions and duration. The expected planning count is not a substitute for inspecting the export.
Use the exercise for cinematic previsualization
A scene study can help explore a camera idea before committing to a finished sequence. Keep the intended movement and landmark checks with the creative brief so you can explain why a take is usable. Avoid presenting a generated camera move as proof that a real location can be filmed or navigated in the same way.
If you need a native 3D asset, use the separate SAM 3D workflow guide and inspect its asset format. If you need a scene-led creative project, explore Clout AI Cinema. Clout publishes this guide and is a separate service; no WorldPlay integration or common-input quality comparison is implied.
Keep the release and result together
For a separate connected-tool production workflow, use the AI video agent brief builder to describe a task and keep job review distinct from completion.
The repository’s release notes also link to a newer HY-World project. A newer project name does not silently update the configuration covered here. Save the exact revision, pipeline, settings, pose and source reference with any output you produce.
This guide checked revision 1588e1336e842b03b0a7860c654ebd7c46bb065e. We executed only the independent local planning calculator and its valid and invalid input checks. No model inference, hardware benchmark or geometric-consistency measurement was performed.
