WorldFoundry
English中文

HunyuanWorld-Mirror

Method overview: Overview of WorldMirror . Our framework employs Multi-modal Tokenization to encode all inputs (images, optional priors including intrinsics, camera poses, and depth maps) into a unified token sequence. The merged.

HunyuanWorld-Mirror from Tencent (Hunyuan): three dimension; Verified.

Edit this pageGitHubMarkdown

HunyuanWorld-Mirror is Tencent's universal 3D reconstruction model: feed it images and it returns point clouds, Gaussian splats, depth, normals, and camera parameters in one pass. In this catalog it is verified — GPU inference with local tencent/HunyuanWorld-Mirror assets produced point-cloud PLY, Gaussian-splat PLY, depth maps, normal maps, and a scene.rrd.

Method overview: Overview of WorldMirror . Our framework employs Multi-modal Tokenization to encode all inputs (images, optional priors including intrinsics, camera poses, and depth maps) into a unified token sequence. The merged.
Method overview: Overview of WorldMirror . Our framework employs Multi-modal Tokenization to encode all inputs (images, optional priors including intrinsics, camera poses, and depth maps) into a unified token sequence. The merged.

The recorded paper is arXiv 2510.10726.

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

Build your run

Pipeline
HunyuanMirrorPipelineworldfoundry.pipelines.hunyuan_world.pipeline_hunyuan_mirror
TaskDefault default
Environmentworldfoundry-unified-cu128
DeviceCUDA 12.8
Runnerworldfoundry.pipeline
Loadingfrom_pretrained
Invocationnative_or_standard_invocation
worldfoundry-eval run \
  hunyuanworld-mirror \
  --pipeline.task-profile default \
  --json

What it is for

Pick hunyuanworld-mirror 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.

WorldFoundry routes inference through the in-tree HunyuanMirrorPipeline and vendored inference-only WorldMirror model code; the official repository serves as provenance for parity checks, not a runtime dependency. Third-party packages (onnxruntime, pycolmap, open3d, gsplat) are environment dependencies, not catalog blockers. It is published by Tencent (Hunyuan).

Catalog tasks include three dimension. Catalog aliases include hunyuan-mirror and hy-world-mirror.

Runtime contract

Optional input_path, confidence_percentile, edge_normal_threshold, edge_depth_threshold, and apply_confidence_mask; artifacts hunyuan_mirror_outputs.zip.

CLI task profiles: default. WorldFoundry binds HunyuanMirrorPipeline.

Family status: Verified.

HunyuanMirrorPipeline accepts image, input, images, or image_path payloads from the standard pipeline runner and writes reconstruction artifacts to the sample output directory. Real inference needs the configured Python environment, local checkpoints, and GPU-compatible torch/gsplat.

WorldFoundry exposes framework inference through the in-tree HunyuanMirrorPipeline and vendored inference-only WorldMirror model code.

Dependencies mirror the official HunyuanWorld-Mirror requirements. gsplat may need a CUDA/PyTorch-specific wheel matching the local torch build.

Running in WorldFoundry

Launch worldfoundry-eval run hunyuanworld-mirror --pipeline.task-profile default 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/HunyuanWorld-Mirror. Prepare the environment and check local assets with bash scripts/inference/prepare_model_infer.sh hunyuanworld-mirror. Add --download to fetch missing public checkpoints.

Status: Verified.