Open Source
Explore the latest AI open-source projects from GitHub and HuggingFace.
Explore the latest AI open-source projects from GitHub and HuggingFace.
HunyuanWorld-Mirror is Tencent Hunyuan's universal feed-forward 3D reconstruction model, accepted to ICML 2026, that produces an unusually complete bundle of 3D outputs — point clouds, multi-view depth maps, camera parameters, surface normals, and 3D Gaussian Splats — from a single forward pass over a video or set of images. Where most prior open-source 3D reconstruction projects either specialize in one output type (depth from monocular video, or pose from structure-from-motion) or require iterative optimization, WorldMirror collapses the full geometric prediction stack into a single inference call and lets the user inject any subset of known priors — camera poses, calibrated intrinsics, depth maps — to sharpen the result. ## The Multi-Modal Prior Prompting Idea The core architectural contribution is what the team calls **Multi-Modal Prior Prompting**: a mechanism that takes whichever geometric priors the user actually has (commonly intrinsics and pose for calibrated rigs, depth for RGB-D capture, or nothing at all for in-the-wild video) and encodes them as conditioning tokens through lightweight encoders. The reported performance numbers show measurable accuracy gains every time another prior is added, which matters in practice because real-world capture pipelines are rarely fully uncalibrated and rarely fully calibrated — they sit somewhere in between, and tools that can use whatever calibration exists without breaking when something is missing are the ones that get adopted. The mechanism is exposed in inference code through three flags — `[camera_pose, depth, intrinsics]` — that turn each prior on or off independently. ## Universal Geometric Prediction Head The second half of the architecture is a unified geometric prediction module that handles the whole spectrum of 3D reconstruction sub-tasks: camera pose estimation, depth estimation, point map regression, surface normal estimation, and novel view synthesis via 3D Gaussian Splatting. Each output type is evaluated separately in the paper — point cloud reconstruction, surface normal estimation, novel view synthesis, depth estimation, and camera pose estimation each get their own benchmark — and the model is reported to be competitive across the full set rather than excelling on one and degrading the rest. The result is a tool that can serve as the geometry layer for downstream pipelines as different as 3D asset creation, robotics navigation, AR scene reconstruction, and view synthesis. ## Outputs That Plug Into Existing Tooling The model emits 3D Gaussian Splats directly, which means the output drops straight into the Gaussian Splatting renderers and editors that have become standard in the neural rendering community over the last two years, and it includes an optional post-3DGS optimization step for users who want to push quality further. The point cloud and depth outputs are conventional enough to feed into Meshroom, COLMAP-style pipelines, or any downstream surface reconstruction tool. By emitting both a feed-forward result and an optional refinement loop, WorldMirror covers both the 'fast preview' and the 'final-quality' use cases without forcing the user to install a different model for each. ## A Family of World Models, Not a One-Shot Release WorldMirror sits inside a deliberately layered Hunyuan world-model lineup. HunyuanWorld-1.0 was the first open-source simulation-capable immersive 3D world generation model in July 2025, HunyuanWorld-1.5 added real-time creation and play in December 2025, HY-World-2.0 and WorldMirror-2.0 landed in April 2026 as the team's state-of-the-art 3D world model, and the original WorldMirror release in October 2025 — the repository this card covers — was upgraded with full training and evaluation code in November 2025. The result is one of the most active open-source 3D research stacks of 2026, and the cadence is fast enough that the November 2025 → ICML 2026 acceptance window is unusual for a paper-track contribution. ## Practical Setup and Footprint The install path is straightforward but assumes CUDA 12.4 and a real GPU: a conda environment, PyTorch 2.4 with cu124 wheels, the project's `requirements.txt`, and `gsplat` for Gaussian Splatting rendering. A Hugging Face Space hosts an online demo for users who want to try the model without local install, and the Gradio-based local demo (`python app.py`) is the recommended starting point. The model weights are auto-downloaded by the inference scripts from `tencent/HunyuanWorld-Mirror` on the Hugging Face Hub. The repository's project page (3d-models.hunyuan.tencent.com/world/) and technical report give the deeper architectural details. ## Where It Fits For practitioners, HunyuanWorld-Mirror is the right pick when a project needs *several* different 3D outputs from the same inputs and wants to avoid maintaining a stack of single-purpose models. It is also the right pick when capture conditions are inconsistent — sometimes calibrated, sometimes not — and a model that gracefully exploits whatever priors happen to exist is more useful than one that demands a fixed input schema. The trade-off against more specialized models is that for any single task, a purpose-built monocular depth model or a dedicated SfM pipeline may still edge out WorldMirror's number on that one metric. The case for WorldMirror is operational: one model, one inference call, one set of weights to maintain, and a clear upgrade path to WorldMirror-2.0 inside HY-World-2.0 when richer world-model capabilities are needed.
graphdeco-inria
Original reference implementation of 3D Gaussian Splatting for real-time radiance field rendering
ahujasid
Connect Blender to Claude AI via MCP for natural-language-driven 3D scene creation and manipulation.
Tencent Hunyuan
Tencent's open-source 3D asset generation system with 13k+ GitHub stars, creating high-resolution textured 3D models from a single image using a two-stage diffusion pipeline.