TripoSR

TripoSR is the official open-source codebase jointly developed by Tripo AI and Stability AI, providing an advanced feed-forward 3D reconstruction model that can quickly generate 3D objects from a single image. The model draws on the principles of the Large Reconstruction Model (LRM) and achieves significant improvements in both speed and reconstruction quality, taking less than 0.5 seconds to process a single input and generate a high-quality 3D model on an NVIDIA A100 GPU. According to the official description, TripoSR outperforms other open-source alternatives in both qualitative and quantitative evaluations on multiple public datasets, and details of the model architecture, training process, and comparisons can be found in its technical report.

Main capabilities:

• Fast feed-forward 3D reconstruction from a single image.
• Generating a high-quality 3D model on an NVIDIA A100 GPU takes less than 0.5 seconds.
• Supports outputting textures rather than only vertex colors, which can be enabled with the --bake-texture option, and --texture-resolution can be used to specify the pixel resolution of the output texture.
• Provides a manual inference script and a local Gradio app.
• Released under the MIT license, including source code, pretrained models, and an interactive online demo.

Usage:

1. If available, install CUDA.
2. Install PyTorch according to the platform (https://pytorch.org/get-started/locally/），需确保本地安装的 The CUDA major version should match the CUDA major version bundled with PyTorch. For example, if CUDA 11.x is installed, then PyTorch compiled with CUDA 11.x must be installed.
3. Update setuptools with pip install --upgrade setuptools.
4. Install other dependencies with pip install -r requirements.txt.

Manual inference: After running the script, the reconstructed 3D model will be saved to the output/ directory. Multiple image paths can be specified, separated by spaces. Under the default options, a single image input requires about 6GB of VRAM. If you want to output textures rather than vertex colors, use the --bake-texture option; you can also use --texture-resolution to specify the pixel resolution of the output texture. Detailed usage of the script can be viewed with python run.py --help.

Local Gradio app: The official method for running the local Gradio app is provided (specific steps are subject to the repository instructions).

Applicable scenarios: This model is aimed at researchers, developers, and creators, and is intended to push the boundaries of 3D generative AI and 3D content creation. It can be used in research and development scenarios for quickly reconstructing 3D objects from a single image.

Limitations and notes:

• Under the default configuration, a single image input requires about 6GB of VRAM.
• CUDA and PyTorch versions need to be correctly configured to match, otherwise problems may occur.
• If AttributeError: module 'torchmcubes_module' has no attribute 'mcubes_cuda' appears, or it indicates that torchmcubes was not compiled with CUDA support and uses the CPU version, this is usually because torchmcubes was not compiled with CUDA support. You should ensure that the local CUDA major version matches the CUDA major version bundled with PyTorch, and ensure setuptools>=49.6.0 (otherwise upgrade via pip install --upgrade setuptools), then reinstall torchmcubes.
• Other specific platform support, performance metrics, or pricing information are not currently provided.