Splatworks — video → splat sequence (Depth Anything)

1Frames

Record a slow orbit around a subject, or pick a video / photos. Video is sliced into evenly-spaced frames, keeping the sharpest moment of each slice.

Video file or photos

Frames to take from a video 12

Slider goes to 240; type in the box for up to 2000.

2Depth & splat options

Depth Anything model (placed next to this page)

Point detail (pixel stride)

Field of view (focal estimate) 55°

Normal-oriented disks (better off-axis)

Advanced

Disk thinness 0.25

Splat opacity 0.92

Each frame is converted feed-forward from its own depth map (no training — that needs multiple views, which is the Trainer stage). This is the SHARP-style per-frame approach.

Depth shaping live · drag to see the render change

show tips

Depth scale a 0.9

Inverse depth is 1/z = b + a·d. Larger a = deeper scene.

Depth shift b 0.12

Larger b pulls everything closer and compresses the far field.

Depth curve (gamma) 1

<1 lifts distant detail, >1 pushes it away.

Relief exaggeration 1

Scales depth about the subject, along each viewing ray.

Trim far (drop disparity below) 0.03

Removes sky / far background.

Trim near (drop disparity above) 1.01

Removes anything too close to the camera.

Edge tolerance 3

Lower culls more across depth steps (kills fg→bg curtains).

Grazing cull 0.18

Higher drops more near-tangent surface (silhouette streaks).

Splat size 1.4

Multiplier on each gaussian’s footprint.

Edge-aware depth cleanup

Median + joint bilateral: snaps depth edges to colour edges. ~100 ms/frame.

no sequence yet

Speed 8 fps Splat size

Output size Load sequence

Generate or load a sequence to see output sizes.

Every frame is always kept — a flipbook with holes is not a flipbook — so the only size knob is how densely each frame is sampled. Save sequence writes a .swseq holding the depth maps, which is the slow part: reload it and the shaping sliders are live again immediately, with no depth inference.

3Progress

Pick a video or photos to begin.

Activity log 0 lines follow

The preview is a real splat rasteriser: each gaussian's 3D covariance is projected to screen and drawn as an oriented, alpha-blended ellipse, depth-sorted front-to-back — the same maths splaTV uses, so preview and export match. Splats seen edge-on appear as thin streaks, not round dots. Above 220,000 splats the preview thins for interactivity; exports always use every gaussian.