--- license: other license_name: nvidia-oneway-noncommercial license_link: https://github.com/nv-dvl/vgg-ttt/blob/main/LICENSE.txt library_name: vggttt pipeline_tag: image-to-3d tags: - vggt - 3d-reconstruction - pointmap - depth-estimation - camera-pose base_model: facebook/VGGT-1B ---

VGG-T³: Offline Feed-Forward 3D Reconstruction at Scale

arXiv Project Page GitHub **[NVIDIA](https://www.nvidia.com/)**     **[University of Toronto](https://www.utoronto.ca/)**     **[Vector Institute](https://vectorinstitute.ai/)** [Sven Elflein](https://selflein.github.io/), [Ruilong Li](https://www.liruilong.cn/), [Sérgio Agostinho](https://sergioagostinho.com/), [Zan Gojcic](https://zgojcic.github.io/), [Laura Leal-Taixé](https://dvl.in.tum.de/team/lealtaixe/), [Qunjie Zhou](https://research.nvidia.com/labs/dvl/author/qunjie-zhou/), [Aljosa Osep](https://aljosaosep.github.io/)
# Model Overview ### Description VGG-T³ reconstructs 3D geometry and camera parameters from image collections and videos. The model scales linearly with the number of input images, which significantly accelerates reconstruction when dealing with a large number of images or long videos. This model is for research and development only. ### License/Terms of Use This model is released under the [NVIDIA OneWay Noncommercial License](https://github.com/nv-dvl/vgg-ttt/blob/main/LICENSE.txt). It may only be used for non-commercial research or educational purposes. ### Deployment Geography Global ### Use Case Primary users: - Computer Vision Researchers: For benchmarking 3D reconstruction and developing neural rendering pipelines. - Augmented Reality/Virtual Reality & Robotics Engineers: For implementing real-time simultaneous localization and mapping (SLAM), scene understanding, and navigation systems. - 3D Content Creators: For rapid conversion of video/images into 3D assets. Primary Use Cases - 3D Reconstruction: Fast, feed-forward estimation of sparse 3D geometry and camera poses from unposed images or video without iterative optimization. - Structure-from-motion (SfM) Replacement: Accelerating initialization for 3D Gaussian Splatting and Neural Radiance Field (NeRF) training by replacing slow SfM (e.g., COLMAP) steps. - Robotic Perception: Instant inference of "pointmaps" (pixel-to-3D coordinates) for manipulation and spatial awareness. ### Release Date GitHub 05/25/2026 via https://github.com/nv-dvl/vgg-ttt
Hugging Face 05/25/2026 via https://huggingface.co/nvidia/vgg-ttt
## References - [Pre-print](https://arxiv.org/abs/2602.23361) - Base model: [VGGT: Visual Geometry Grounded Transformer](https://arxiv.org/abs/2503.11651) ## Quick Start Install the [code](https://github.com/nv-dvl/vgg-ttt) (preferably in a conda environment): ```bash git clone https://github.com/nv-dvl/vgg-ttt.git && cd vgg-ttt pip install torch==2.7.1 torchvision==0.22.1 --index-url https://download.pytorch.org/whl/cu126 pip install -r requirements.txt ``` VGG-T³ is compatible with the VGGT API and can be used in a similar way: ```python from vggttt.nets.vggt.models.vggt import VGGT from vggttt.nets.vggt.img import load_and_preprocess_images vggttt = VGGT.from_pretrained("nvidia/vgg-ttt").eval().cuda() image_names = ["path/to/imageA.png", "path/to/imageB.png", "path/to/imageC.png"] images = load_and_preprocess_images(image_names).to("cuda") preds = vggttt.infer(images) # Dict containing the predicted outputs with the following keys: # - 'pose': [#images, 4, 4] Camera-to-world transformation # - 'intrinsics': [#images, 3, 3] Pinhole camera matrix # - 'pts3d': [#images, height, width, 3] Per-pixel points in world coordinates # - 'conf': [#images, height, width] Per-pixel confidence in range ]1, inf[ # - 'depth': [#images, height, width, 1] Per-pixel depth ``` ## Model Architecture **Architecture Type:** Transformer
**Network Architecture:** Vision-Transformer (ViT)
**This model was developed based on:** [VGGT](https://arxiv.org/abs/2503.11651)
**Number of model parameters:** 1.19 × 10⁹
## Input(s) **Input Type(s):** Image, video
**Input Format(s):** - Image: Red, Green, Blue (RGB)
- Video: .mov, .mp4
**Input Parameters:** - Video: Three-Dimensional (3D)
- Collection of images: Three-Dimensional (3D)
**Other Properties Related to Input:** - Video is converted to images at frame rate of the video. - Maximum resolution per image: 518x518 ## Output(s) **Output Type(s):** - Camera parameters (extrinsics, intrinsics) - Point cloud (3D point per pixel) **Output Format(s):** - Point cloud: X, Y, Z - Camera intrinsics: Focal length - Camera extrinsics: Rotation matrix, translation vector **Output Parameters:** - Point cloud: Three-Dimensional (3D)
- Camera intrinsics: Two-Dimensional (2D)
- Camera extrinsics: Three-Dimensional (3D)
**Other Properties Related to Output:** - One 3D point is predicted per input pixel in each image. - One set of camera parameters (intrinsics + extrinsics) is predicted per image. Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA's hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.
## Software Integration **Runtime Engine(s):** N/A **Supported Hardware Microarchitecture Compatibility:**
* NVIDIA Ampere
* NVIDIA Blackwell
* NVIDIA Hopper
* NVIDIA Volta
**Preferred/Supported Operating System(s):** Linux The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.
## Model Version(s) - v1.0: Initial release ## Training, Testing, and Evaluation Datasets ### Training Datasets #### **DynamicReplica (DynamicStereo)** **Link:** https://dynamic-stereo.github.io/
**Data Modality**: Image, Depth, Camera extrinsics and intrinsics. **Image Training Data Size** [Less than a Million Images]
**Data Collection Method by dataset** [Synthetic]
**Labeling Method by dataset** [Synthetic]
**Properties:** 145,200 stereo frames (from 524 videos). The dataset contains synthetic rendered videos of scanned environments with dynamic objects (people/animals). It includes stereo images, disparity maps, and camera parameters.
----- #### **Hypersim** **Link:** https://github.com/apple/ml-hypersim
**Data Modality**: Image, Depth, Camera extrinsics and intrinsics. **Image Training Data Size** [Less than a Million Images]
**Data Collection Method by dataset** [Synthetic]
**Labeling Method by dataset** [Synthetic]
**Properties:** \~77,400 images. Photorealistic synthetic dataset for indoor scene understanding. Contains RGB, depth, semantic segmentation, and camera parameters.
----- #### **OmniData** **Link:** https://github.com/EPFL-VILAB/omnidata/tree/main/omnidata_tools/dataset#readme
**Data Modality**: Image, Depth, Camera extrinsics and intrinsics. **Image Training Data Size** [1 Million to 1 Billion Images]
**Data Collection Method by dataset** [Hybrid: Automatic/Sensors, Synthetic]
**Labeling Method by dataset** [Synthetic]
**Properties:** \~14 million rendered images from various meshes that are partially obtained from real world scans.
----- #### **CubifyAnything** **Link:** https://github.com/apple/ml-cubifyanything
**Data Modality**: Image, Depth, Camera extrinsics and intrinsics. **Image Training Data Size** [1 Million to 1 Billion Images]
**Data Collection Method by dataset** [Automatic/Sensors]
**Labeling Method by dataset** [Hybrid: Human, Automatic/Sensors]
**Properties:** The dataset aligns handheld RGB-D captures with laser scans, creating a large-scale labeled dataset for 3D object detection.
----- #### **ScanNet++** **Link:** https://scannetpp.mlsg.cit.tum.de/scannetpp/
**Data Modality**: Image, Depth, Camera extrinsics and intrinsics. **Image Training Data Size** [1 Million to 1 Billion Images] **Data Collection Method by dataset** [Human]
**Labeling Method by dataset** [Hybrid: Human, Automatic/Sensors]
**Properties:** Millions of frames (video) / \~460 scenes. High-resolution real-world 3D indoor semantic scene understanding dataset using laser scanning and DSLR/iPhone capture.
----- #### **Mapillary Metropolis** **Link:** https://www.mapillary.com/dataset/metropolis
**Data Modality**: Image, Depth, Camera extrinsics and intrinsics. **Image Training Data Size** [Less than a Million Images]
**Data Collection Method by dataset** [Automatic/Sensors]
**Labeling Method by dataset** [Hybrid: Human, Automatic/Sensors]
**Properties:** \~28,000 images. City-scale dataset containing street-level panoramic images, LiDAR point clouds, and aerial imagery.
----- #### **MatrixCity** **Link:** https://city-super.github.io/matrixcity/
**Data Modality**: Image, Depth, Camera extrinsics and intrinsics. **Image Training Data Size** [Less than a Million Images]
**Data Collection Method by dataset** [Synthetic]
**Labeling Method by dataset** [Synthetic]
**Properties:** \~67,000 images (Small City block). Large-scale synthetic city dataset for neural rendering, generated using Unreal Engine 5.
----- #### **MegaDepth** **Link:** https://www.cs.cornell.edu/projects/megadepth/
**Data Modality**: Image, Depth, Camera extrinsics and intrinsics. **Image Training Data Size** [Less than a Million Images]
**Data Collection Method by dataset** [Automated]
**Labeling Method by dataset** [Automatic/Sensors]
**Properties:** \~150,000 reconstructed images. Derived from Internet photo collections (Flickr) using Structure-from-Motion (SfM) and Multi-View Stereo (MVS) to generate depth maps.
----- #### **Mid-Air** **Link:** https://midair.ulg.ac.be/
**Data Modality**: Image, Depth, Camera extrinsics and intrinsics. **Image Training Data Size** [Less than a Million Images]
**Data Collection Method by dataset** [Synthetic]
**Labeling Method by dataset** [Synthetic]
**Properties:** \~420,000 video frames. Synthetic dataset for low-altitude drone flight, featuring multi-modal data (RGB, depth, normals, stereo) across different climate conditions.
----- #### **Mapillary Planet-scale Depth dataset** **Link:** https://www.mapillary.com/dataset/depth
**Data Modality**: Image, Depth, Camera extrinsics and intrinsics. **Image Training Data Size** [Less than a Million Images]
**Data Collection Method by dataset** [Hybrid: Human, Automatic/Sensors]
**Labeling Method by dataset** [Automatic/Sensors]
**Properties:** \~750,000 images. Dense depth dataset derived from street-level imagery using SfM, covering a wide variety of locations and camera models.
----- #### **ParallelDomain4D** **Link:** https://github.com/basilevh/gcd?tab=readme-ov-file#paralleldomain-4d
**Data Modality**: Image, Depth, Camera extrinsics and intrinsics. **Image Training Data Size** [1 Million to 1 Billion Images]
**Data Collection Method by dataset** [Synthetic]
**Labeling Method by dataset** [Synthetic]
**Properties:** Synthetic video dataset rendered using the Parallel Domain engine. Designed for dynamic novel view synthesis and 4D reconstruction.
----- #### **Virtual KITTI 2** **Link:** https://europe.naverlabs.com/proxy-virtual-worlds-vkitti-2/
**Data Modality**: Image, Depth, Camera extrinsics and intrinsics. **Image Training Data Size** [Less than a Million Images]
**Data Collection Method by dataset** [Synthetic]
**Labeling Method by dataset** [Synthetic]
**Properties:** Photorealistic synthetic video dataset mimicking the real KITTI dataset, used for object detection, tracking, and depth estimation.
----- #### **Waymo Open Dataset** **Link:** https://waymo.com/open/
**Data Modality**: Image, Depth, Camera extrinsics and intrinsics. **Image Training Data Size** [1 Million to 1 Billion Images]
**Data Collection Method by dataset** [Hybrid: Human, Automatic/Sensors]
**Labeling Method by dataset** [Hybrid: Human for semantic annotations, Automatic/Sensors for depth]
**Properties:** Large-scale autonomous driving dataset collected with high-resolution cameras and LiDAR.
----- #### **Common Objects in 3D (CO3Dv2)** **Link:** https://ai.meta.com/datasets/co3d-dataset/
**Data Modality**: Image, Depth, Camera extrinsics and intrinsics. **Image Training Data Size** [1 Million to 1 Billion Images]
**Data Collection Method by dataset** [Human]
**Labeling Method by dataset** [Automatic/Sensors]
**Properties:** \~6 million images (v2 version of the dataset). Large-scale dataset of object-centric videos captured with smartphones, annotated with camera poses and point clouds via structure-from-motion.
----- #### **Kubric** **Link:** https://github.com/google-research/kubric?tab=readme-ov-file
**Data Modality**: Image, Depth, Camera extrinsics and intrinsics. **Image Training Data Size** [Less than a Million Images]
**Data Collection Method by dataset** [Synthetic]
**Labeling Method by dataset** [Synthetic]
**Properties:** A data generation pipeline using PyBullet and Blender to create photo-realistic synthetic videos with rich annotations. We generate a synthetic dataset using based on assets from [Google Scanned Objects](https://research.google/blog/scanned-objects-by-google-research-a-dataset-of-3d-scanned-common-household-items/), environments from [HDRI Haven](https://polyhaven.com/hdris)
----- #### **Wild RGB-D** **Link:** https://github.com/wildrgbd/wildrgbd
**Data Modality**: Image, Depth, Camera extrinsics and intrinsics. **Image Training Data Size** [1 Million to 1 Billion Images]
**Data Collection Method by dataset** [Human]
**Labeling Method by dataset** [Automatic/Sensors]
**Properties:** Large-scale dataset comprising \~20,000 RGB-D videos of objects captured in the wild using iPhones. Includes camera poses and object masks.
----- #### **BlendedMVS** **Link:** https://github.com/YoYo000/BlendedMVS
**Data Modality**: Image, Depth, Camera extrinsics and intrinsics. **Image Training Data Size** [Less than a Million Images]
**Data Collection Method by dataset** [Synthetic]
**Labeling Method by dataset** [Synthetic]
**Properties:** Multi-view stereo dataset. 3D models are reconstructed from real images, then rendered to create synthetic training pairs blended with the original images.
----- #### **DL3DV-10K** **Link:** https://dl3dv-10k.github.io/DL3DV-10K/
**Data Modality**: Image, Camera extrinsics and intrinsics. **Image Training Data Size** [1 Million to 1 Billion Images]
**Data Collection Method by dataset** [Human]
**Labeling Method by dataset** [Automatic/Sensors]
**Properties:** Large-scale scene dataset containing over 10,000 videos of diverse real-world scenes captured for Deep Learning based 3D Vision. Camera poses are obtained using structure-from-motion.
----- #### **Spring Benchmark** **Link:** https://spring-benchmark.org/
**Data Modality**: Image, depth, Camera extrinsics and intrinsics. **Image Training Data Size** [Less than a Million Images]
**Data Collection Method by dataset** [Synthetic]
**Labeling Method by dataset** [Synthetic]
**Properties:** High-resolution photo-realistic computer-generated imagery (Blender) for optical flow and stereo benchmarking.
----- #### **TartanAir** **Link:** https://theairlab.org/tartanair-dataset/
**Data Modality**: Image, Depth, Camera extrinsics and intrinsics. **Image Training Data Size** [1 Million to 1 Billion Images]
**Data Collection Method by dataset** [Synthetic]
**Labeling Method by dataset** [Synthetic]
**Properties:** Challenging visual SLAM dataset collected in photorealistic simulation environments with various lighting/weather conditions.
----- #### **UnrealStereo4K** **Link:** https://github.com/fabiotosi92/SMD-Nets#unrealstereo4k
**Data Modality**: Image, Depth, Camera extrinsics and intrinsics. **Image Training Data Size** [Less than a Million Images]
**Data Collection Method by dataset** [Synthetic]
**Labeling Method by dataset** [Synthetic]
**Properties:** High-resolution synthetic stereo dataset generated using Unreal Engine 4.
### Evaluation Datasets #### **7scenes** **Link:** [7scenes](https://www.microsoft.com/en-us/research/project/rgb-d-dataset-7-scenes/)
**Data Modality**: Image, depth, Camera extrinsics and intrinsics. **Image Training Data Size** [Less than a Million Images]
**Data Collection Method by dataset** [Human]
**Labeling Method by dataset** [Automatic/Sensors]
**Properties:** 7-Scenes is a RGB-D dataset featuring seven indoor scenes captured with a Kinect camera. The dataset includes RGB images (640x480), depth maps, and camera-to-world poses.
--- #### **DTU MVS Dataset** **Link:** https://roboimagedata.compute.dtu.dk/?page_id=36
**Data Modality**: Image, depth, Camera extrinsics and intrinsics. **Image Training Data Size** [Less than a Million Images]
**Data Collection Method by dataset** [Human]
**Labeling Method by dataset** [Automatic/Sensors]
**Properties:** An industrial robot arm was mounted with a structured light scanner. This allowed for structured light scans corresponding to each image in the data set. The images were taken by one of the cameras in the structured light scanner.
--- #### **ETH3D Stereo Benchmark** **Link:** https://www.eth3d.net/overview **Data Modality**: Image, depth, Camera extrinsics and intrinsics. **Image Training Data Size** [1 Million to 1 Billion Images]
**Data Collection Method by dataset** [Human]
**Labeling Method by dataset** [Automatic/Sensors]
**Properties:** The multi-view stereo / 3D reconstruction benchmark covers a variety of indoor and outdoor scenes. Ground truth geometry has been obtained using a high-precision laser scanner. A DSLR camera as well as a synchronized multi-camera rig with varying field-of-view was used to capture images.
--- #### **ScanNet** **Link:** http://www.scan-net.org/ **Data Modality**: Image, depth, Camera extrinsics and intrinsics. **Image Training Data Size** [Less than a Million Images]
**Data Collection Method by dataset** [Human]
**Labeling Method by dataset** [Hybrid: Automatic/Sensors, Human]
**Properties:** ScanNet is an RGB-D video dataset containing 2.5 million views in more than 1500 scans, annotated with 3D camera poses, surface reconstructions, and instance-level semantic segmentations. To collect this data, we designed an easy-to-use and scalable RGB-D capture system that includes automated surface reconstruction and crowdsourced semantic annotation.
--- #### **Sintel** **Link:** http://sintel.is.tue.mpg.de/ **Data Modality**: Image, depth, Camera extrinsics and intrinsics. **Image Training Data Size** [Less than a Million Images]
**Data Collection Method by dataset** [Synthetic]
**Labeling Method by dataset** [Synthetic]
**Properties:** 1,064 frames organized into 23 video sequences. The data is Computer-Generated Imagery (CGI) derived from the open-source 3D animated short film Sintel.
--- #### **TUM RGB-D SLAM dataset** **Link:** https://cvg.cit.tum.de/data/datasets/rgbd-dataset **Data Modality**: Image, depth, Camera extrinsics and intrinsics. **Image Training Data Size** [Less than a Million Images]
**Data Collection Method by dataset** [Human]
**Labeling Method by dataset** [Automatic/Sensors]
**Properties:** Large dataset containing RGB-D data and ground-truth data with the goal to establish a novel benchmark for the evaluation of visual odometry and visual SLAM systems. The dataset contains the color and depth images of a Microsoft Kinect sensor along the ground-truth trajectory of the sensor. The data was recorded at full frame rate (30 Hz) and sensor resolution (640x480). The ground-truth trajectory was obtained from a high-accuracy motion-capture system with eight high-speed tracking cameras (100 Hz). --- #### **Bonn RGB-D Dynamic Dataset** **Link:** https://www.ipb.uni-bonn.de/data/rgbd-dynamic-dataset/index.html **Data Modality**: Image, depth, Camera extrinsics and intrinsics. **Image Training Data Size** [Less than a Million Images]
**Data Collection Method by dataset** [Human]
**Labeling Method by dataset** [Automatic/Sensors]
**Properties:** This is a dataset for RGB-D SLAM, containing dynamic sequences. It provides 24 dynamic sequences, where people perform different tasks, such as manipulating boxes or playing with balloons, plus 2 static sequences. For each scene we provide the ground truth pose of the sensor, recorded with an Optitrack Prime 13 motion capture system. The sequences are in the same format as the TUM RGB-D Dataset, so that the same evaluation tools can be used. Furthermore, they provide a ground truth 3D point cloud of the static environment recorded using a Leica BLK360 terrestrial laser scanner. --- #### **KITTI Vision Benchmark Suite** **Link:** https://www.cvlibs.net/datasets/kitti/ **Data Modality**: Image, depth, Camera extrinsics and intrinsics. **Image Training Data Size** [Less than a Million Images]
**Data Collection Method by dataset** [Human]
**Labeling Method by dataset** [Hybrid: Automatic/Sensors, Human]
**Properties:** Challenging real-world computer vision benchmarks. Tasks of interest are: stereo, optical flow, visual odometry, 3D object detection and 3D tracking. For this purpose, they equipped a standard station wagon with two high-resolution color and grayscale video cameras. Accurate ground truth is provided by a Velodyne laser scanner and a GPS localization system. The datasets are captured by driving around the mid-size city of Karlsruhe, in rural areas and on highways. Up to 15 cars and 30 pedestrians are visible per image. --- #### **Wayspots** **Link:** https://nianticlabs.github.io/ace/wayspots.html **Data Modality**: Image, depth, Camera extrinsics and intrinsics. **Image Training Data Size** [Less than a Million Images]
**Data Collection Method by dataset** [Human]
**Labeling Method by dataset** [Automatic/Sensors]
**Properties:** 10 small outdoor scenes stemming from Niantic's Map-Free dataset. Each scene was scanned twice, independently by two users. Posed mapping images come from real-time visual odometry on the user's phones. Depth or 3D point clouds are not available. Evaluation ground truth comes from SfM. # Inference **Acceleration Engine:** N/A
**Test Hardware:** NVIDIA A100
## Ethical Considerations NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).
## Citation If you find this work useful, please cite: ```bibtex @inproceedings{elflein2026vggttt, title = {VGG-T\textsuperscript{3}: Offline Feed-Forward 3D Reconstruction at Scale}, author = {Elflein, Sven and Li, Ruilong and Agostinho, S{\'e}rgio and Gojcic, Zan and Leal-Taix{\'e}, Laura and Zhou, Qunjie and Osep, Aljosa}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, year = {2026} } ```