Image-GS: Content-Adaptive Image Representation via 2D Gaussians

Upload an image and configure parameters to train a 2D Gaussian Splatting representation.

Configuration

Input Image

Basic Parameters

Experiment Name

Name for this training run

Number of Gaussians

Number of Gaussians (for compression rate control). More = higher quality but slower training

100 50000

Maximum Training Steps

Maximum number of optimization steps. Default: 10000

100 20000

Quantization

Enable Quantization

Enable bit precision control of Gaussian parameters. Reduces memory usage.

Position Bits

Bit precision of individual coordinate dimension

4 32

Scale Bits

Bit precision of individual scale dimension

4 32

Rotation Bits

Bit precision of Gaussian orientation angle

4 32

Feature Bits

Bit precision of individual feature dimension

4 32

Initialization

Initialization Mode

Gaussian position initialization mode. Gradient uses image gradients, saliency uses attention maps.

gradient saliency random

Random Ratio

Ratio of Gaussians with randomly initialized position (default: 0.3)

0 1

Visualization

Visualize Gaussians

Visualize Gaussians during optimization (default: True)

Save Image Every N Steps

Frequency of rendering intermediate results during optimization (default: 100)

200 10000

Status

Training Progress

Training Logs

Initialization Map

Current Training Results

Current Render

Gaussian ID

Browse Training Steps

Slide to view results from different training steps (disabled during training)

0 10000

Final Results

Final Render

Download Final Checkpoint (.pt)

Image-GS: Content-Adaptive Image Representation via 2D Gaussians

Configuration

Basic Parameters

Quantization

Initialization

Loss Weights

Learning Rates

Optimization

Visualization

Training Progress

Initialization Map

Current Training Results

Final Results