IGGT4D¶

Development line: project:iggt4d · thread iggt4d
Last event: 2026-07-22 · 1 dated since 2026-07-22 · Researched: 2026-09-05 · confidence: medium

What it is¶

IGGT4D processes video frames sequentially to predict camera, geometry, persistent object identities, and instance masks together.

• Streaming 4D reconstruction
• Pose estimation
• Object tracking
• Open-vocabulary segmentation

The 1B parameter checkpoint requires accepting access terms on Hugging Face; dataset, benchmark, training, and evaluation code are not released yet. It works for inference experiments on custom image sequences, but not for reproducible training or benchmarking.

Development line¶

• 2026-07-22 — IGGT4D GitHub Pages site was linked. Causal streaming model that unifies scene geometry and persistent object identities.

What changed¶

2026-07-22 — IGGT4D was presented as Streaming 4D Instance-Grounded Geometry Transformer: a causal streaming model that unifies scene geometry and persistent object identities.

How to use this¶

From 2026-07-22, practitioners should treat the linked IGGT4D GitHub Pages site as a project reference point, while verifying its specific guidance or release status before relying on it.

1. Clone the official repository and create a Conda environment with Python 3.10. — https://github.com/HorizonRobotics/IGGT4D
2. Install PyTorch; the README gives the example of PyTorch 2.8.0 with CUDA 12.8, then install the package with pip install -e .. — https://github.com/HorizonRobotics/IGGT4D
3. Accept the checkpoint access terms on Hugging Face, download it, and save it as checkpoints/model.safetensors. — https://huggingface.co/HorizonRobotics/IGGT4D
4. Run python streaming_infer.py for the demo, or pass --image-dir and --output-dir for a custom sequence of RGB frames. — https://github.com/HorizonRobotics/IGGT4D
5. If needed, enable --camera for external camera parameters and save results via --save-npy or --save-ply. — https://github.com/HorizonRobotics/IGGT4D

Best practices¶

• Keep streaming mode as the default for long sequences; --mode full is the offline full-attention variant. — https://github.com/HorizonRobotics/IGGT4D
• Do not plan training, evaluation, or comparison on the official benchmark before release: the repository lists them as pending artifacts. — https://github.com/HorizonRobotics/IGGT4D

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• For the 2026-07-22 event, the primary arXiv source specifies the paper submission date as 2026-07-21 16:00:01 UTC; this confirms the title, ten authors, and the InsScene4D-147K scale, but gives no separate dated publication for 2026-07-22.
• The exact date when inference code and the checkpoint became available is not stated on the primary pages read; therefore we did not add it as a separate dated event.
• The public project page states that the dataset and benchmark will arrive later, and the repository also awaits training and evaluation code.

Sources¶

Agent brief¶

• Subject: project:iggt4d, thread iggt4d, 1 dated events 2026-07-22 → 2026-07-22.
• Practical note: From 2026-07-22, practitioners should treat the linked IGGT4D GitHub Pages site as a project reference point, while verifying its specific guidance or release status before relying on it.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.