OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
Image & Video AI

Traceanything

TraceAnything is an open-source Python and PyTorch computer-vision model that maps pixels across frames to continuous 3D trajectories, representing visual sequences as trajectory fields. Its repository provides inference scripts and an interactive 3D viewer for exploring results.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Traceanything — OpenViST project profile in Image & Video AI.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use TraceAnything when you need local inference that converts ordered video frames or image pairs into pixel-level trajectory fields and related outputs for analysis or reconstruction.

Who it is for

  • Computer-vision researchers analyzing motion across video frames or paired views
  • Developers incorporating trajectory estimates into 4D reconstruction workflows
  • Practitioners with a local Python, PyTorch and high-memory GPU setup

Documented uses

  • Estimate continuous 3D trajectories for pixels across ordered video frames.
  • Create trajectory-field representations from image pairs.
  • Inspect control points, confidence maps, predicted times and foreground masks from model output.
  • Use trajectory estimates as input to downstream 4D reconstruction workflows.
Before you choose

Requirements & limits

The public repository is identified as Apache-2.0 licensed. The documented workflow is a local Python/PyTorch setup with downloaded model weights.

  • The supplied inference workflow requires images ordered in time; unordered image-set support is described as future work.
  • The provided examples are tested on a single GPU with at least 48 GB VRAM, and CUDA 12.8 is the stated tested configuration.
  • The inference script automatically resizes and crops images, and downsamples scenes with more than 40 views.
  • Optional point-guided masks use a separate SAM2 installation and checkpoint.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Run inference on ordered scene images

Generate trajectory-field outputs for scenes stored as ordered images using the repository's inference script.

Before you start

  • A prepared TraceAnything environment and checkpoint at checkpoints/trace_anything.pt
  • Scene images organized in subdirectories under an input directory
  • Images ordered in time, such as video frames or paired images

Steps

  1. Place each scene's images in its own subdirectory under examples/input, or specify another input directory.

  2. Run python scripts/infer.py, setting --input_dir, --output_dir and --ckpt flags if using non-default paths.

  3. Inspect each scene's output.pt file and the generated masks under the output directory.

Expected outcome

Each scene receives an output.pt containing 3D control points, confidence maps, foreground masks and predicted time values; mask images are also saved.

Guide limitations

  • The script assumes images are ordered in time; support for unordered inputs is described as future work.
  • Images are resized with a 512-pixel long side and cropped to a multiple of 16; scenes above 40 views are automatically downsampled.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
15 October 2026
Original website title
ByteDance-Seed/TraceAnything · Official homepage title · retained in its source language
How we review sourcesSuggest a correction