App reference
3D reconstruction reference: photogrammetry and 3DGS
26.10.1110
Novaber · Content reviewed
On this page
Turn multi-view photos or video into 3D point clouds and Gaussian splats on Windows. Combine COLMAP photogrammetry, NeRF, and 3D Gaussian Splatting with foreground selection, keyframe previews, stage progress, and result exports for object, plant, and scene digitization.
Capabilities and input / output
-
Create tasks from photos or video and select sharp frames at the requested count
-
Choose SAM3 text prompts or positive and negative point annotations
-
Combine annotations from multiple keyframes, preview all frames, then confirm for reconstruction
-
Inspect each point-guided preview and the current final foreground masks and transparent RGBA results
-
Estimate camera poses and build sparse or MVS dense point clouds with COLMAP
-
Train and export NeRF / nerfacto and 3DGS / splatfacto results
-
Browse point clouds and Gaussian results in the app and download engine-specific outputs
-
Track stage parameters, progress, logs, and versions with cancellation and failure retry
Workflow
1. Import media
Create a task and select MVS, NeRF, or 3DGS. Video uses the configured sharp-frame count; images retain their input sequence.
2. Foreground processing (optional)
Text prompts segment automatically. Review and confirm point-based previews; video supports propagation across keyframes.
3. Cameras and spatial structure
Match multi-view features, estimate camera poses, and build a sparse cloud for MVS, NeRF, and 3DGS.
4. Reconstruct with selected engines
Run the selected dense-cloud, neural-field, or Gaussian reconstruction. Image scaling follows task parameters, with stage status, logs, and version history available.
5. Inspect and export
Inspect foreground masks, RGBA, and 3D outputs separately. Browse point clouds or Gaussian results in the app and download files produced by the selected engines.
Requirements
- GPU: NVIDIA GPU Required
- VRAM: 12GB+ VRAM
- Images and models: Image >10GB
Review your first result
These checks are suggestions based on published inputs and outputs. Use a small sample to verify processing, preview and export together; no measured accuracy or speed is provided here.
Identify MVS, NeRF or 3DGS before reviewing its result. If using foreground processing, inspect masks, transparent RGBA and the 3D result separately. A foreground preview does not verify camera estimation or reconstruction. Confirm downloaded files belong to the chosen task and engine.
Version, limitations and license
Current catalog version: 26.10.1110.
Third-party names and trademarks belong to their respective owners. Use and distribution are governed by each project's license and do not imply endorsement of Novaber.
Product details and installation
Choose an engine by the intended 3D result
| Intended result | Route | Inspect separately |
|---|---|---|
| Camera poses and sparse structure | COLMAP | Camera estimation and sparse cloud used by later stages |
| Dense point cloud | MVS | Dense reconstruction after camera estimation |
| Neural radiance field | NeRF / nerfacto | Engine-specific training state and output |
| Gaussian scene | 3DGS / splatfacto | Gaussian output, in-app preview and download |
Choose the deliverable before creating the task. Foreground images are not point clouds or Gaussian scenes. These engine options come from the public catalog; select actual parameters in the current app.
Operate and review each stage
1. Import multi-view photos or video
Create a task and choose MVS, NeRF or 3DGS. Images retain their input sequence; video selects sharp frames at the requested count. Start with a small set covering the intended views, keep originals and record engine and scaling settings.
2. Prepare foreground when needed
Foreground processing is optional. Text prompts and point confirmation differ: the text route continues through its workflow, while point-based previews need review and confirmation. Video can combine annotations from multiple keyframes. Review all frames before committing.
Inspect subject masks and transparent RGBA at this stage, including point-guided previews and the final foreground. Confirm the intended inputs enter reconstruction. Foreground success does not establish camera or reconstruction success.
3. Check spatial structure and the selected engine
Inspect COLMAP camera poses and sparse structure, followed by the MVS, NeRF or 3DGS stage. Use task progress, logs and history to locate failure; retain settings and errors before cancelling or retrying.
4. Preview and download matching outputs
Browse point clouds or Gaussian results in the app and confirm downloads belong to the task and engine. Review masks, RGBA and 3D outputs separately. Identify the missing stage when reporting a problem. See reconstruction questions.
Prepare the environment
For NVIDIA GPU workflows, check driver major >= 575 and Windows driver >= 576.57 for the project’s CUDA 12.9.1 baseline. See the driver and CUDA baseline for version comparison, unknown results and container checks. Check the selected app’s GPU model, VRAM and model requirements separately.