co-tracker vs vggt
Video point tracking and 3D scene reconstruction compared
co-tracker follows selected or grid-sampled points through video, with offline and chunk-by-chunk inference. vggt estimates cameras, depth, point maps, and tracks from one or more images, combining several 3D geometry predictions in one model.
co-tracker: Track Points Across Video Frames
CoTracker tracks selected or grid-sampled points through video using a transformer-based model. It offers offline and online inference, pretrained checkpoints, and tools for evaluation and training.
vggt: Reconstruct 3D Scenes from Images
VGGT is a feed-forward vision model that estimates cameras, depth, point maps, and point tracks from one or more images. It suits researchers and developers who need fast 3D scene reconstruction without a traditional multi-stage pipeline.
| co-tracker | vggt | |
|---|---|---|
| Language | Jupyter Notebook | Python |
| License | NOASSERTION | NOASSERTION |
| Stars | 5.1k | 14.5k |
| Forks | 392 | 1.6k |
| Last analyzed | Oct 3, 2026 | Oct 3, 2026 |
Key differences
- co-tracker focuses on point trajectories across video; vggt focuses on scene reconstruction and predicts camera parameters, depth, and 3D point maps as well as tracks.
- co-tracker supports offline and online chunk-by-chunk video inference; vggt supports single-view and multi-view inputs and provides APIs and demos for images or video.
- co-tracker is listed as Jupyter Notebook, while vggt is listed as Python.
- Both report a NOASSERTION repository license, but co-tracker notes that most of the project is CC-BY-NC with some portions under separate terms; vggt distinguishes its non-commercial original checkpoint from a commercial checkpoint with separate terms.
- co-tracker includes evaluation workflows for TAP-Vid and Dynamic Replica; vggt supports COLMAP-format export, optional bundle adjustment, and Gaussian splatting workflows.
- Both are unarchived and include training code; co-tracker documents a 32-GPU Kubric training run, while vggt notes that its pretrained model is large and practical inference is best suited to a CUDA-capable GPU.
Choose co-tracker if you…
- need trajectories for selected or grid-sampled video points.
- want to process video offline or in streaming chunks.
- need supplied point-tracking evaluation workflows or training code for Kubric and pseudo-label fine-tuning.
Choose vggt if you…
- need camera, depth, and 3D point-map estimates from one or more images.
- want to export reconstructions in COLMAP format for downstream geometry or Gaussian splatting workflows.
- want to explore zero-shot single-view reconstruction or fine-tune on a custom dataset.
Compare the numbers side by side, or add more projects
This comparison is generated with AI from the OSRepos analyses of both projects. Always check each project's repository and documentation before choosing.