GeometryCrafter
| Property | Value |
|---|---|
| Developer | TencentARC |
| Model Type | Geometry Estimation |
| Hardware Requirements | 40GB GPU (full resolution) / 22GB GPU (low resolution) |
| Repository | https://github.com/TencentARC/GeometryCrafter |
What is GeometryCrafter?
GeometryCrafter is an innovative AI model developed by TencentARC that estimates temporally consistent, high-quality point maps from open-world videos. This groundbreaking technology enables advanced 3D/4D reconstruction and depth-based video editing capabilities, making it a valuable tool for computer vision applications.
Implementation Details
The model operates at various performance levels depending on configuration: 1.27FPS for full resolution (1024x576) processing and up to 2.49FPS for low-resolution (384x640) processing. It employs diffusion priors to ensure consistent geometry estimation across video frames.
- Supports multiple resolution modes for different hardware capabilities
- Includes both standard and deterministic variant implementations
- Features comprehensive visualization tools via Viser
- Provides extensive dataset evaluation capabilities
Core Capabilities
- High-quality point map generation from videos
- Temporal consistency maintenance across frames
- Scale-invariant point map estimation
- Affine-invariant depth estimation
- Support for various video resolutions
Frequently Asked Questions
Q: What makes this model unique?
GeometryCrafter stands out for its ability to maintain temporal consistency while processing open-world videos, offering both standard and deterministic variants for different use cases. Its diffusion priors enable robust geometry estimation across diverse scenarios.
Q: What are the recommended use cases?
The model is ideal for 3D/4D reconstruction tasks, depth-based video editing, and geometry estimation in open-world scenarios. It's particularly useful for applications requiring consistent point map generation from video content.
