How Many Photos for a Gaussian Splat Do You Actually Need?

How many photos for a gaussian splat is the wrong question to lead with, even though it's the first one everyone asks. The honest answer is a range that moves with the subject: a few dozen stills for a single object, somewhere between 300 and 600 for a typical room, thousands for a full floor or site. None of those numbers do much on their own.

What actually decides the outcome is overlap, how much of each photo repeats what the one before it saw. A reconstruction engine finds a point in 3D space by matching it across two or more views. Photos with gaps between them leave the software nothing to match. A smaller set shot with deliberate overlap reconstructs cleanly where a larger, careless one does not. Each photo becomes raw material for a splat's points, not a mesh, and what a 3D gaussian splat actually is covers that mechanism if the term itself is new.

Count is a proxy for coverage. Overlap is the mechanism that makes the count matter at all.

How many photos for a gaussian splat do you actually need?

For a single object, a few dozen frames shot in a steady arc around it. For one room, 300 to 600 photos, or 5 to 10 minutes of steady video instead. For a full floor, building or site, plan on thousands of frames, usually gathered as continuous video across several passes rather than counted stills.

Those are starting points, not targets to hit exactly. A smaller space with simple geometry, four flat walls and no clutter, can reconstruct well under the low end of its range. A room full of glass, mirrors, or repeating tile patterns often needs more, because there is less unique detail in any one frame for the software to lock onto. The subject decides the floor. The surfaces decide how far past it you should go.

How much should each photo overlap the next?

Captures Studio's own capture guide puts the overlap target at 60 to 70 percent between consecutive photos. That means most of what appears in one frame should still be visible, from a slightly different angle, in the next one. Walk too fast between shots, or turn too far, and that shared area shrinks until the software has nothing to connect the two views with.

Coverage works the same way across height and angle, not just position. A single pass at eye level captures walls but misses the ceiling and skips most of the floor. Three passes cover a room properly: one around chest height, one overhead angled down (with roughly 30 to 60 percent of the frame taken up by floor), and one low, near waist height, angled up. Anything you want reconstructed in detail (a fireplace, a piece of furniture, a countertop) gets its own short arc at a few different heights.

Where more photos actually helps

Reflective and repetitive surfaces need the extra volume. Glass, polished tile, and a plain white wall all give a reconstruction engine less to distinguish one frame from the next, so denser overlap compensates for the missing detail. The same is true of fine texture at close range, a product with visible grain or stitching benefits from more frames shot a short distance away rather than fewer from further back.

Scale is the other real reason to shoot more. A multi-room interior or a full site needs volume simply to reach every surface once, not because any single wall benefits from extra passes. That's a coverage problem dressed up as a count problem.

Where more photos stops helping

Past the coverage target, extra frames from the same angle and distance add processing time without adding information. The tutorial line is blunt about it: prioritise full coverage over hitting an exact number.

Movement in the scene hurts more than a thin frame count does. A gaussian splat assumes the room holds still between shots, so a person walking through frame, a pet, or curtains moving in the wind reconstructs as a blur or a ghost rather than resolving into anything sharper with more photos. Inconsistent lighting across a long session (direct sun moving across a wall between the start and end of a shoot) causes the same kind of failure. More frames of a moving or shifting scene just means more evidence of the problem, not less.

What a first capture session should look like

Clear the room first. No people walking through, no pets, nothing swaying, and lighting that stays roughly the same from the first photo to the last. Direct sunlight through a window changes fast enough to cause trouble across a ten-minute session.

Shoot the three-pass pattern: chest height around the whole room, an overhead pass angled down toward the floor, and a low pass angled up toward the ceiling. Add a tight arc around anything you specifically want to look good, a mantel, a kitchen island, a staircase. A tripod or a stabiliser and a slightly higher frame rate cut down on the motion blur that a handheld pass at dusk tends to introduce.

If this is a first attempt, video is the more forgiving format to start with. Captures Studio treats an ordinary MP4 or MOV clip the same way it treats a folder of stills once processing starts, so a slow, steady walkthrough removes the guesswork of deciding how many individual shots to take. Once the overlap habit is there, moving to dedicated stills for a harder subject (a small object, or a highly reflective room) is a smaller jump. The full three-pass method, including the arcs around focal objects, is laid out step by step in the capture tutorial.

FAQ

How many photos do I need for a small object?

A few dozen is usually enough for something under about a metre across, shot in a steady arc that keeps consecutive frames overlapping through the whole turn. Lighting matters more here than count. Keep it even and diffuse, since a shiny surface needs more coverage to reconstruct cleanly than a matte one does.

Can I use video instead of taking individual photos?

Yes. A steady 5 to 10 minute walkthrough covers a typical room, and something smaller can work with a minute or two, as long as the pass stays slow and doesn't jump around. The footage gets sampled into frames during processing, so the same overlap rules apply either way.

What happens if I take too few photos?

Gaps in coverage show up as holes, blur, or a flattened patch where the software had nothing to triangulate from. The fix is rarely more photos from the same spot. It's a short second pass aimed at the specific angle or wall the first pass missed.

Does more overlap always improve the result?

Only up to a point. Sixty to seventy percent between consecutive frames gives the reconstruction enough shared detail without spending a session on near-duplicate shots. Past that, extra frames mostly add processing time rather than sharper surfaces.

How many photos does a full building or site need?

Large sites routinely run into the thousands of frames, usually gathered as continuous video across several passes instead of counted stills. At that scale the count stops being the useful number. Coverage per zone (has every room had its three passes) is what to track instead.

Should I capture a room that has people or pets moving through it?

No. A gaussian splat assumes the scene stays still between shots, so anything that moves between frames (a contractor walking through, a dog wandering into a doorway, blinds swinging in an open window) reconstructs as a blur or a ghost rather than a solid surface. Clear the room and keep it static for the whole pass.

The number worth writing down before a shoot isn't 300 or 600. It's 60 to 70 percent, checked as you go rather than counted at the end. A session that hits that overlap on every wall, corner and ceiling will beat a much larger one that skipped a corner because nobody walked back to cover it.

That's the difference between a splat that comes back clean on the first processing run and one that comes back with a hole where the hallway used to be.