Multi-view scene reconstruction
Infer consistent geometry and camera information from overlapping photographs or extracted video frames.
Independent tool overview
HunyuanWorld-Mirror at a glance
HunyuanWorld-Mirror is Tencent's source-available feed-forward model for reconstructing scene geometry, cameras, depth, normals, point clouds, and 3D Gaussians from image sequences or video frames.
Visit the official HunyuanWorld-Mirror site ↗
- Developer
- Tencent Hunyuan
- Product type
- Feed-forward 3D reconstruction model
- Primary inputs
- Image sequences or video frames
- Optional priors
- Camera poses, intrinsics, and depth maps
- Core outputs
- Points, depth, normals, cameras, and 3D Gaussians
- Recommended stack
- Python 3.10, PyTorch 2.4, CUDA 12.4
- Publication
- Accepted to ICML 2026
- Current successor
- WorldMirror 2.0 in HY-World 2.0
- License
- Tencent community license with regional and use limits
- Last reviewed
- August 29, 2026
Overview
What HunyuanWorld-Mirror is
HunyuanWorld-Mirror, also called WorldMirror, is a research-oriented 3D reconstruction model from Tencent Hunyuan. It takes one or more images, or frames extracted from a video, and predicts a collection of geometric representations in one forward pass.
Despite some early descriptions, WorldMirror is not itself a text-to-3D world generator. Its job is multi-view reconstruction: estimating point clouds, per-view depth, surface normals, camera poses and intrinsics, and 3D Gaussian parameters. Tencent's broader HunyuanWorld systems use reconstruction and generation components together.
The model can run without calibrated inputs or accept any available combination of camera poses, camera intrinsics, and depth maps as geometric priors. Tencent's published evaluations show those priors can materially improve reconstruction and novel-view results.
The original repository remains useful and was accepted to ICML 2026, but it now points users to WorldMirror 2.0 inside HY-World 2.0 for the newer world-composition pipeline. The original release includes inference, training, evaluation, a Gradio demo, model weights, Gaussian-splatting optimization, and COLMAP-oriented exports.
The code and weights are publicly downloadable but use Tencent's HunyuanWorld-Mirror Community License, not a permissive open-source license. The license excludes the European Union, United Kingdom, and South Korea, adds distribution and acceptable-use duties, restricts using outputs to improve other AI models, and requires separate permission for certain products above one million monthly active users.
Use cases
Who HunyuanWorld-Mirror is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
3D Gaussian initialization
Produce Gaussian-splatting parameters and optionally refine them with the included gsplat-based optimization workflow.
Camera and depth estimation
Predict camera poses, intrinsics, per-view depth, and confidence maps when calibration data is incomplete.
Novel-view research
Render and evaluate views beyond the original camera positions using the reconstructed 3D representation.
Geometry research
Study a unified model that handles point maps, depth, normals, cameras, and splats through shared features.
Custom training experiments
Use the released training and evaluation code to fine-tune selected prediction heads on compatible datasets.
Capabilities
Core HunyuanWorld-Mirror features
Multi-modal prior prompting
Encodes any available camera poses, calibrated intrinsics, and depth maps as structured conditioning tokens.
Universal geometric prediction
Uses one architecture for point reconstruction, depth, normals, camera estimation, and novel-view synthesis.
Single-pass inference
Predicts several scene representations together instead of running an independent pipeline for every geometry task.
Point clouds with confidence
Returns world-coordinate 3D points and per-point confidence for each input view.
Depth and surface normals
Produces camera-frame depth and normal maps with confidence values.
Camera recovery
Estimates camera-to-world poses, intrinsic matrices, translation, rotation, and fields of view.
3D Gaussian outputs
Predicts Gaussian means, opacity, scale, rotation, and spherical-harmonic color features.
Novel-view rendering
Uses the reconstructed Gaussians to render viewpoints not present in the input sequence.
COLMAP-compatible export
Saves cameras, images, point data, Gaussian PLY files, and other initialization artifacts for downstream workflows.
Optional splat optimization
Refines the feed-forward result through the included gsplat training example when extra processing time is acceptable.
Training and evaluation code
Includes configurable prediction heads and benchmark workflows for point maps, normals, views, depth, and camera poses.
Process
How the HunyuanWorld-Mirror workflow works
Step 1
Choose the correct generation
Use the original WorldMirror for reproducible reconstruction research or evaluate WorldMirror 2.0 when building a new HY-World pipeline.
Step 2
Review the license first
Confirm the deployment region, user scale, distribution plan, training use, and acceptable-use obligations before downloading or integrating the model.
Step 3
Capture overlapping views
Provide an ordered image sequence or a video with enough scene coverage and parallax to support geometric reconstruction.
Step 4
Prepare the recommended environment
Set up Python 3.10, the documented PyTorch and CUDA versions, repository dependencies, and gsplat packages on compatible hardware.
Step 5
Add trustworthy priors when available
Supply calibrated intrinsics, camera poses, or depth maps and set the matching conditioning flags rather than passing uncertain metadata.
Step 6
Run feed-forward inference
Process the views at the configured resolution and save point, depth, normal, camera, and Gaussian predictions with confidence outputs.
Step 7
Inspect geometry before refinement
Visualize points, normals, depth, camera trajectories, and novel views to catch inconsistent coverage, drift, or low-confidence regions.
Step 8
Optimize and export if needed
Run the optional 3D Gaussian optimization, then export the Gaussian and COLMAP artifacts required by the downstream renderer or editor.
Step 9
Validate on the target task
Measure reconstruction accuracy, view quality, runtime, memory, and failure cases on the scenes and cameras the application will actually use.
Cost
HunyuanWorld-Mirror pricing and free plan
Tencent does not charge a download fee for the public code and weights, but access is governed by a restrictive community license. Compute, storage, engineering, and any additional commercial license are separate costs.
Research and permitted use
No download fee
Use the public repository and checkpoint only where the community license grants rights.
- Source code, inference weights, training code, and evaluation code are public
- License territory excludes the EU, UK, and South Korea
- Acceptable-use and distribution conditions apply
- Outputs may not be used to improve another AI model except WorldMirror derivatives
Large-scale commercial use
Permission required
A separate Tencent license is required in the user-scale case defined by the community agreement.
- Applies when the licensee's covered products or services exceeded 1 million monthly active users in the preceding month at release
- Approval is discretionary
- Company and intended-use information must be submitted
- Terms and price are not publicly listed
Self-hosted compute
Infrastructure cost
Run inference, optimization, training, or evaluation on infrastructure you provide.
- GPU, CPU, memory, storage, and electricity costs apply
- Long image sequences and Gaussian optimization increase resource use
- Teams own setup, security, uptime, and maintenance
- Training datasets have their own access and license terms
Pricing checked . Check current pricing at the source ↗
Assessment
HunyuanWorld-Mirror strengths and limitations
Where it stands out
- Predicts several useful 3D representations through one shared feed-forward model
- Works with images or video frames without requiring every geometric prior
- Can incorporate camera, intrinsics, and depth priors when better calibration is available
- Outputs confidence values that help identify uncertain geometry
- 3D Gaussian and COLMAP-oriented exports fit established reconstruction workflows
- Optional optimization provides a quality-versus-speed choice after fast initialization
- Inference, training, evaluation, demo code, and model weights are all publicly available
- Acceptance to ICML 2026 and published benchmark protocols make the research easier to examine
What to consider
- WorldMirror reconstructs from visual observations; it is not a direct text-to-3D world generator
- The original release has been followed by WorldMirror 2.0 inside HY-World 2.0
- The Tencent community license is not a permissive open-source license and excludes several major regions
- Large-scale commercial products may need separate written permission from Tencent
- The license restricts using model outputs to improve unrelated AI models
- CUDA, PyTorch, gsplat, model checkpoints, and 3D processing dependencies make local setup technical
- Reconstruction quality depends on input coverage, scene content, calibration, and the reliability of supplied priors
- Feed-forward results may still need Gaussian optimization and cleanup before production rendering or asset use
Compare
HunyuanWorld-Mirror alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Content Creator
World Labs Marble
A hosted, creator-oriented system that generates persistent 3D worlds from text, image, or video with less local setup.
Explore World Labs Marble →Miscellaneous
Sam 3D
Meta's models focus on reconstructing individual objects and people from a single image rather than full multi-view scenes.
Explore Sam 3D →Design
Seed 3D
A simulation-oriented image-to-3D model aimed at producing structured assets for physical and embodied environments.
Explore Seed 3D →Questions
HunyuanWorld-Mirror FAQs
What is HunyuanWorld-Mirror?
HunyuanWorld-Mirror is Tencent's feed-forward model for reconstructing scene geometry and cameras from images or video frames. It predicts point clouds, depth, normals, camera parameters, and 3D Gaussians.
Can WorldMirror create a 3D world from text?
Not by itself. WorldMirror reconstructs geometry from visual inputs. Tencent's broader HunyuanWorld pipelines include separate generation, navigation, stereo, and composition components for text- or image-conditioned worlds.
What inputs does WorldMirror accept?
It accepts a sequence of images or frames extracted from video. Camera poses, camera intrinsics, and depth maps are optional priors that can be supplied in any available combination.
What does WorldMirror output?
The model can output world-coordinate points, depth maps, surface normals, confidence maps, camera poses and intrinsics, and 3D Gaussian parameters for rendering and downstream optimization.
Is HunyuanWorld-Mirror open source?
The code and weights are publicly available, but they use Tencent's HunyuanWorld-Mirror Community License with geographic, scale, distribution, acceptable-use, and model-improvement restrictions. It is better described as source-available than permissively open source.
Can HunyuanWorld-Mirror be used in the EU, UK, or South Korea?
The published community license says it does not apply in those regions and treats use outside its defined territory as unauthorized. Obtain qualified legal advice and a suitable license before use there.
What happened after the original WorldMirror release?
Tencent released WorldMirror 2.0 as a component of HY-World 2.0 in April 2026. The original repository remains available and was accepted to ICML 2026, but new projects should compare the successor.
How much does HunyuanWorld-Mirror cost?
There is no public download fee for permitted use of the repository and weights. Users pay their own compute and engineering costs, and some large-scale commercial deployments require separate permission from Tencent.
Bottom line
Our HunyuanWorld-Mirror verdict
HunyuanWorld-Mirror is valuable for researchers and technical 3D teams that need a unified, inspectable reconstruction model with rich geometry outputs and optional priors. It should not be presented as a one-click text-to-world product. New projects should compare WorldMirror 2.0, and every organization should resolve the community license's regional and commercial restrictions before implementation.
Visit HunyuanWorld-Mirror website ↗Similar Miscellaneous Tools
View all Miscellaneous AI tools →
PokeeResearch 7B
PokeeResearch 7B- Pokee AI's SOTA, open-source deep research agent
Mistral Studio
Mistral AI Studio - A platform for companies to move from AI prototypes to production with built-in tools for performance tracking, testing, and security.
Fish Audio S1
Fish Audio S1 -
Rnj-1
Rnj-1 - Essential AI's open-source 8B parameter model that matches or outperforms similar-sized rivals
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.