◈ Splat Lab BEHEERVerbinden…
◈

Breng een scène in beeld.

Start met een voorbereide foto- of videoset. Het model krijgt stap voor stap meer detail; pauzeer zodra je de GPU nodig hebt.

NEW TRAINING RUN

Build a new model

    Choose your prepared capture

    This starts a separate training family. To improve a saved version, use Improve model on that version.

    Prepared datasets include the images and solved camera positions.

    Choose a starting quality

    More detail as training progresses
    ¼ resolution→½ resolution→Full resolution

    The same model continues learning at each stage. You can pause and resume at any point.

    Advanced settings

    Downscale factor : starting step. A factor of 1 uses original resolution.

    Review your training plan

    After creation, Start training begins camera loading, staged training, export and quality evaluation. The live journey shows each stage.

    Creating a model saves its settings. It won’t use the GPU until you press Start.

    IMPROVE A SAVED MODEL

    Refine regions

      Where should detail improve?

      Let Splat Lab work through the site one area at a time. You only need the map when you want a specific area.

      Inspect or customize the area map

      How much refinement?

      Adaptive detail uses image gradients and a bounded growth budget. Both methods choose update strength on selection photos, then check separate audit photos once.

      Advanced settings

      More steps give each area more optimization time; they do not guarantee more detail. Adaptive detail can add at most 100,000 splats per area and 1,000,000 across this run, within the active limit. Conservative refinement keeps the count unchanged.

      Review your refinement journey

      1. Check visible photo coverage for each area.
      2. Train one area at a time within the memory budget.
      3. Compare held-out views and boundaries; retain source geometry if an update fails.
      4. Save the composed version in this training family.

      Areas without sufficient coverage or within-limit splats are skipped with a reason. Memory pressure pauses the run safely. Existing versions remain available.

      Review and create, then press Start training. Grouping versions never merges unrelated coordinate systems.

      SHARPER DETAIL · TILE BY TILE

      Train in tiles

        Which tiles?

        Each tile trains on its own, from scratch, with every photo that sees it. Tiles share the survey coordinates, so they fit together without alignment and form one 3D model and one 2D map.

        Survey points seen from above

        Creating saves the plan. Training starts when you press Start.

        2D MAP · COORDINATES

        Reference point

        One point is enough: scale and north come from the drone positions, the point fixes the remaining 1–2 m offset. With two or more points the rotation is fitted as well and residuals are shown.

        Organize this model

        Moving a training family brings its checkpoints and refinement branches together. A shared group is for organization; it does not establish coordinate alignment.

        KEEP THE ORIGINAL · SAVE A NEW VERSION

        Clean up details

        Remove faint splats or overly large splats. Inspect the result in 3D and compare the original before choosing a preferred version; filters can also remove thin real structures.

        Size is the largest Gaussian standard deviation in survey-coordinate units. Leave it blank to keep all sizes.

        Crop by survey coordinates

        Start cleanup runs two bounded scans on the CPU. The source model and its appearance coefficients remain available.

        ADD CAPTURE → CHECK CAMERAS → REFINE

        Add photos and video

        Add overlapping photos or video of this scene. We retain the originals, select sharp video frames, and check new cameras against the existing survey. Your current model remains available.

        Choose files to upload.

        Measured lens calibration (optional)

        Use a profile for the exact camera, lens, focus, zoom, stabilization and resolution of these files. Use separate batches for different cameras. Without a profile, registration estimates the new camera calibration.

        Frames cover the full video; the sharpest candidate in each time interval is kept. Camera registration can reject frames without enough overlap. Once prepared, Improve model trains a new refinement using the added views.

        Completed uploads survive a reload. Start preparation from the revision to see each processing stage.

        PRINT → CAPTURE → CHECK → USE

        Calibrate your lens

        A checkerboard can measure lens distortion. One picture cannot reliably determine all lens parameters. Capture 15–25 sharp views with the board at different angles and positions, including the image edges.

        Download printable A4 PDFDownload vector SVG

        10 × 7 squares, 9 × 6 inner corners, 20 mm squares. Print at 100% / Actual size on paper or matte sticker stock. Mount flat on rigid backing; five squares must measure 100 mm. A bent sticker gives a misleading calibration. Keep the full board visible. For video, use full-resolution still frames from the same recording mode. This measures residual distortion in your saved images, including any camera correction already applied.