VGGT-1B-Point on Replicate turns one or more photos—or a video—into 3D point data, with an optional downloadable GLB point cloud. The easiest first run is in Replicate’s playground: upload images of the same scene, keep the initial settings, run the model, then inspect the result in a compatible 3D viewer. It is a convenient way to try a reconstruction without setting up a local GPU, but the output is not a guaranteed measurement or polished 3D object.
What VGGT-1B-Point does
VGGT stands for Visual Geometry Grounded Transformer. The broader VGGT project is a feed-forward model that infers 3D scene attributes from one or more views. The VGGT-1B-Point listing on Replicate focuses on point maps and omits the depth head, which its description says is redundant with the point-map head. Do not assume this point variant returns the same depth output as the separate full vufinder/vggt-1b model.
As an Amazon Associate I earn from qualifying purchases.
A point map represents scene locations as 3D points. A point cloud is a collection of such points, rather than a complete surface mesh or finished, textured object. A GLB is a 3D file format that compatible viewers can open. The hosted model can return an optional GLB point cloud and JSON data for each image.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
The JSON includes world points and confidence values, pose encoding, image information and original image dimensions; a mask may also be included. These are model outputs to inspect, not proof that the reconstruction is accurate.
#1 Best Overall
What to prepare before your first run
Choose images or a video of one scene
Use a small group of clear views that show the same subject or scene. Different viewpoints can give the model more visual information, though the listing does not prescribe an ideal number of images or promise a particular result. You can also submit a video.
Check the supported formats
The listing shows support for JPG, JPEG, PNG and WEBP images, plus MP4, AVI and MOV video. Replicate says it pads images and sampled video frames to a shared aspect ratio and resizes them to a maximum dimension of 518 pixels.
Run VGGT-1B-Point in the Replicate playground
- Open the VGGT-1B-Point model page on Replicate and choose the playground.
- Upload one or more supported images, or a video. Keep the subject visible across the views you provide.
- Start with the listed defaults. The page shows
return_pcdas true,to_base64as true, andsampling_rateas 24. It also lists an alpha-channel blend option. - If you use video,
sampling_ratemeans every nth frame is sampled; the first and last frames are included. Changing it affects how many intermediate frames are processed. The listing does not establish one best setting for every clip. - Run the prediction and inspect the preview. If a GLB point cloud is returned, download it and open it in a compatible 3D viewer. The output also provides URLs for per-image JSON data.
Understand the output and its limits
The GLB point cloud is useful for visually exploring the inferred scene in 3D. It is not, by itself, a watertight mesh, a textured asset or a reliable source of metric-scale measurements. The listing provides no quantitative quality guarantee for this hosted point variant. Results can depend on the input, so check the reconstruction visually rather than treating its confidence values as independent validation.
Meta’s broader VGGT project describes predicting camera parameters, point maps, depth maps and 3D point tracks. That description applies to the broader project, not necessarily to every hosted variant. For this Replicate point model, the listing specifically describes omitting the depth head.
Choose playground, API or local use
| Route | Best for | What to expect |
|---|---|---|
| Replicate playground | A first experiment without writing code | Hosted inference with a variable per-run charge. |
| Replicate API | Repeatable runs or integration into an application | The model page provides Python, Node.js and HTTP API examples. You need an API token, and hosted inference is billed. |
| Local repository workflow | People who want to manage their own environment and compute | Meta’s repository includes Python dependencies and a local quick start; setup and local hardware management are your responsibility. |
For automation, follow the current example on the model page and pin the model version identifier it shows so that your integration names a specific version. Treat the API token as a secret. Replicate’s model page describes invocation, but it does not establish data-retention or token-handling guarantees beyond the service’s applicable terms.
Check the live cost and runtime before running
Replicate’s page showed an estimate of about $0.057 per run (about 17 runs per dollar) and typical completion in 59 seconds on Nvidia L40S hardware when observed on 2026-10-07. These are mutable page estimates, not fixed prices or a runtime guarantee; Replicate says cost and prediction time vary with inputs. Check the live model page before budgeting or starting repeated runs.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Check checkpoint terms for your use
The Replicate listing says it uses VGGT-1B-Commercial weights. Meta’s repository distinguishes the commercial checkpoint, which permits commercial use under its terms, from the original checkpoint, which is non-commercial. Read the precise terms that apply to the checkpoint and service you use before relying on the model for commercial work. The broader VGGT project was published at CVPR 2025; subsequent project updates, including the repository’s VGGT-Omega announcement in May 2026, do not by themselves establish a change to this Replicate endpoint.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
