📄 EDGS: Eliminating Densification for Efficient Convergence of 3DGS (CVPR 2026)
GPU Space recommended. This demo performs 3D Gaussian Splatting optimization and is not intended for CPU-only Spaces. Duplicate this Space to ZeroGPU, L40S, A100, or another 24GB+ CUDA GPU Space for best results. A10G/L4 may work for smaller inputs; T4/CPU are not recommended.
🛠️ How to Use This Demo
- Upload a front-facing video or a folder of images of a static scene.
- Use the sliders to configure the number of reference views, correspondences, and optimization steps.
- First press on preprocess Input to extract frames from video(for videos) and COLMAP frames.
- Then click 🚀 Start Reconstruction to actually launch the reconstruction pipeline.
- Watch the training visualization and explore the 3D model. ‼️ If you see nothing in the 3D model viewer, try rotating or zooming — sometimes the initial camera orientation is off.
✅ Best for scenes with small camera motion. ❗ For full 360° or large-scale scenes, we recommend the Colab version (see project page).
🏋️ Training Visualization
🌐 Interactive 3D Preview
This preview is a browser-friendly RGB point cloud. Download the full Gaussian Splatting .ply below for the real EDGS output.
📦 Output Files
📖 Detailed Overview
If you uploaded a video, it will be automatically cut into a smaller number of frames (default: 16).
The model pipeline:
- 🧠 Runs PyCOLMAP to estimate camera intrinsics & poses (~3–7 seconds for <16 images).
- 🔁 Computes 2D-2D correspondences between views. More correspondences generally improve quality.
- 🔧 Optimizes a 3D Gaussian Splatting model for several steps.
🎥 Training Visualization
You will see a visualization of the entire training process in the "Training Video" pane.
🌀 3D Model
The right-side preview is an interactive Plotly RGB point cloud derived from the generated Gaussian model. Rotate, pan, and zoom directly in the Plotly viewport.