It's all in the capture.
A splat can never outshine its source video — nothing downstream fixes a bad capture. The good news: a great one takes exactly two rules.
Rule · 01 — Camera settings
Freeze every frame.
Every frame must be tack-sharp. Motion blur erases the very pixels MacSplat matches between frames — a soft capture reconstructs soft, forever.
Why — structure from motion
Before any splatting, SfM finds the same visual features in frame after frame and — from how they shift — solves where the camera was and where every point sits in 3D. That point cloud is the skeleton your splat is trained on, and it only trusts sharp, honest frames.
Shutter as fast as light allows.
Grain is recoverable, blur is not — aim for 1/500s or faster. Shooting video? Move like you're carrying a full cup. Shooting photos? Stop walking for every single frame.
Stabilization off. All of it.
EIS hides shake by cropping into the sensor and re-framing every frame — pixels move in ways the camera never did, and the 3D solve warps.
Why it breaks SfM
SfM assumes pixel motion equals camera motion. Stabilization decouples the two, so the solved camera path is a path that never happened — geometry comes out subtly warped, or the solve fails outright. A slightly shaky, honest frame beats a smooth, dishonest one.
Lock exposure. Lock focus.
Auto modes re-decide every frame — brightness pumps, focus hunts, and the same wall becomes three different walls. Tap the subject, lock both, then shoot.
Why — one surface, one color
Training averages what every frame says about a point. If exposure drifts, the same point is bright in one frame and dark in the next, and the splat settles on a muddy compromise — that's the milky haze in bad captures. Autofocus adds its own problem: it quietly changes the lens, so the intrinsics the solve assumed no longer hold.
Rule · 02 — Camera movement
Move. Don't pan.
Structure from motion — so the camera has to actually go somewhere. Keep the lens on the subject and travel: sideways, in, out, around, higher. Rotation is not motion. Standing and swivelling adds no parallax , and without parallax there is no depth.
What is parallax?
As your viewpoint moves, near things sweep past faster than far things. That speed difference is parallax — the only depth cue SfM gets, and it triangulates every 3D point from it. Rotating in place slides everything together: zero parallax, zero depth. Drag the camera:
Your feet make the depth. Your wrist makes nothing.
Panning shows new pixels but zero new viewpoints — flat. Travelling with the lens locked on the subject makes every step a new viewpoint — depth falls out. Hover either side to drive the camera.
Then take the camera somewhere.
Slow is smooth · smooth is sharpOne slow loop if you can get all the way around — subject framed dead-center.
Radial passes are gold: in close for detail, back out for context.
Low, eye level, high — the rings sew the whole scene together.
None of these is the rule — the rule is that the camera keeps moving. A full orbit if the subject allows it, a straight walk down a façade, a figure-eight through a room, a slow rise up a staircase: all of it is motion, and motion is what the solve eats.
Move slowly enough.
MacSplat samples ≈3 frames a secondEach sampled frame sees a slice of the scene. Move slowly and those slices overlap, so every surface is described from several angles — move fast and dark gaps open up: nothing to match, nothing to reconstruct. Whatever path you take, keep about an arm's length between viewpoints.
That's all
Shoot. Drop. Done.
Five checks, then MacSplat takes over — SfM, training, splat. All on your Mac.
THAT'S THE WHOLE SKILL. THE REST IS COMPUTE.
Drop your capture
