Splattr docs

Appearance

Full Settings Reference ​

This page provides a detailed explanation of every option available in the manual job configuration page.

Processing Preview ​

This panel gives a real-time overview of your job's cost.

  • Initial Cost: A small, one-time credit charge for job setup. You can use cloud storage providers to set this to 0
  • Spending Rate: The cost in credits per minute of processing time on the selected server tier.
  • Maximum Cost: The maximum amount of credits you will be charged for the job. In the future the job will get paused if it exceeds this amount.
  • Balance Required: The total credits your account needs to start this job. This is a soft check against your balance and won't deduct the amount from your balance if you have enough.

Quick Presets ​

These are one-click presets that apply common configurations.

  • Fast Processing: Optimizes for the quickest possible result.
  • Balanced Speed & Quality: The recommended default for a great mix of speed, cost, and quality. It is also the default preset option.
  • Maximum Quality: Optimizes for the best possible visual fidelity.

Server Configuration ​

Choose the GPU hardware your job will run on.

  • Standard Server: Recommended for most projects. Powered by elite consumer GPUs (e.g., RTX 5090) with 24GB of VRAM.
  • Heavy Duty Server: For extremely large datasets. Powered by data-center GPUs (e.g., H200) with 141GB of VRAM to prevent out-of-memory errors.

Data Processing ​

Controls how raw media is converted into a 3D scene. This section is skipped for Pre-aligned Data projects.

Basic Settings ​

  • Camera Type: Specify the camera lens used. Perspective for standard phones/DSLRs, Fisheye for fisheye lenses, Equirectangular for 360° cameras.
  • Matching Strategy: The algorithm for finding matching points between images. Sequential is faster and best for videos. Vocab Tree is more robust for unordered photo sets.
  • Number of Downscales: Creates lower-resolution image copies to improve processing. Recommended 1-3 for 4K+ footage.
  • Frame Sampling Rate: Frames to extract per second of video. Recommended: 2-5.

Advanced Settings ​

For Experts

These settings are for advanced users. When in doubt, leave them at their defaults.

  • SFM Tool: hloc is a newer, more advanced tool than the standard colmap but way slower.
  • Feature & Matcher (only if hloc is selected): The specific algorithms for finding keypoints. disk+lightglue is a robust modern option.
  • Multiple camera scene (only if hloc is selected): Enable if you used different cameras or lenses in the same shoot.
  • Same dimentions: A helper toggle assuming all images are the same resolution.
  • Use SFM depth: Uses depth information from the SfM process to guide initial Gaussian placement. Can improve density.

Training ​

Configures the Gaussian Splat generation process.

Basic Settings ​

  • Maximum number of iterations: Total training steps. 30,000 is a strong default. More steps can add detail but have diminishing returns.
  • Image resolution: Train on Full, Half (0.5), or Quarter (0.25) resolution images. Lower resolutions are much faster and use less memory.
  • Training method: The core algorithm. splatfacto is the best balance of speed and quality. -big and -w variants are for more complex scenes.
  • Spherical Harmonics Degree: Controls the complexity of lighting and reflections. 3 is very high quality. 0 is flat lighting but much faster.

Scene and camera settings ​

  • Camera optimizer: Further refines camera poses during training. Can fix slight inaccuracies from the data processing step and reduce blurriness.
  • Initialize randomly: If your initial point cloud is sparse, this adds a large number of random gaussians to help fill in empty space.
  • Background color: random is best for training as it helps define object edges. We do not recommend changing this.

Densification & Geometry ​

Settings controlling how Gaussians are split and removed.

  • Refine Every: How often (in iterations) to run the densification and culling process. More frequent can be better but is slower.
  • Densify Gradient Threshold: The sensitivity for splitting Gaussians. Lower is more sensitive, creating more detail but potentially more noise.
  • Opacity Culling Threshold: Threshold for removing transparent gaussians. Lower values preserve more detail and faint structures but increase the final gaussian count (and size).
  • Stop splitting at: The iteration at which the system stops adding new Gaussians and moves into a refinement-only phase.

Rendering & Quality ​

  • Rasterization mode: Antialiased produces higher quality renders with fewer shimmering artifacts, but is slightly slower than Classic.
  • SSIM loss weight: Balances between L1 loss (pixel-perfect sharpness) and SSIM loss (better perceptual quality). A common value is 0.2.
  • Use scale regularization: A technique to prevent Gaussians from becoming excessively large and "blobby.". Helps reduce "floaters" or large, spiky, and misshapen gaussians that can appear in the scene.