3D catalogue · Verified local path

RecGen local guide

Non-commercial RGB-D object reconstruction model that exports textured meshes, GLB assets and Gaussian splats.

Choose an app

Start with the recommended app, or choose another supported option.

Compare 3D models
Open model filesRecommended verified starting pointOpen advanced setupPython runtime for experienced usersOpen official sourcePublisher repository and primary evidence

Official example

Play the official source preview directly below. The media stays on the official publisher server and is never hosted by LocalClaw.

What it does

Image To 3dMesh ReconstructionGaussian SplattingPose EstimationTexturing

Toyota Research Institute publishes RecGen inference and training code plus public Hugging Face checkpoints for the stereo denoiser, SLAT denoiser, configuration files and TRELLIS decoder/base assets. The official setup uses Linux, CUDA, PyTorch, spconv and diff-gaussian-rasterization, with Pixi environments for CUDA 12.1 or 11.8 and a Docker path for NVIDIA Container Toolkit users. Single-view and multiview inference consume RGB, depth, mask and intrinsics, then write mesh.obj, posed_mesh.obj, overlay.png, metadata.json, optional gaussian.ply and posed_gaussian.ply, and optional textured_mesh.glb when nvdiffrast plus the Gaussian rasterizer are built. No consumer VRAM table is published, so LocalClaw records 64 GB RAM and 24 GB NVIDIA VRAM as a conservative workstation floor for local inference with GLB/splat export headroom.

Hardware figures are practical entry floors, not performance guarantees. Resolution, duration, precision, offloading and runtime versions can materially change memory use.

Strengths

  • Official TRI code and released Hugging Face checkpoints
  • Single-view or multi-view RGB-D reconstruction with 6-DoF pose prediction
  • Exports OBJ meshes, Gaussian PLY splats and optional textured GLB assets

Limits to know

  • Non-commercial TRI license and CC BY-NC checkpoint metadata
  • Requires depth, object mask and camera intrinsics rather than a casual RGB-only prompt
  • GLB and turntable outputs need extra CUDA rasterization dependencies