How to Turn a Video into a Gaussian Splat Online (No GPU)

Tutorial · 2026-07-14 · 8 min read · by SplatMart Team

Upload a raw video and turn it into a downloadable Gaussian splat without extracting frames or owning an NVIDIA GPU. This guide covers capture technique, cloud processing with Leighton, input limits, pricing, cleanup, and export.

The easiest way to make a Gaussian splat from video is no longer to extract hundreds of still frames by hand. Record a slow, overlapping path around a static subject, upload the original clip, and let a cloud GPU solve the camera positions and train the scene. SplatMart's Leighton now accepts raw video in the browser and returns an editable PLY, so the computer doing the upload does not need an NVIDIA GPU.

What video-to-splat processing actually does

A video is a sequence of images. A reconstruction does not need every near-identical frame, so Leighton samples useful viewpoints across the full clip, caps the working set at 3,000 frames, solves where the camera was for each one, and trains millions of 3D Gaussians to reproduce those views. The output is a navigable 3D scene, not a flat video wrapped around a model.

Before filming: choose a scene that can reconstruct

• Keep the subject and environment still. Moving people, pets, traffic, screens, or leaves can become smears and floaters.
• Use soft, consistent light. Avoid switching lights on mid-shot or walking between bright sun and deep shade.
• Choose surfaces with visible detail. Large blank walls, mirrors, transparent glass, and shiny metal are difficult for camera alignment.
• Clean the camera lens and use the normal lens rather than changing zoom during the capture.

How to film a strong capture

Move slowly enough that individual frames stay sharp, but keep moving so the clip contains genuinely different viewpoints. Maintain roughly 70–80% visual overlap as you circle the subject. Capture at eye level, then add a higher and lower pass so the top and underside are not missing. For a room, walk a connected route around the perimeter and point back into the space; for an object, make two or three complete orbits.

• Do not stand still and pan from one spot, Gaussian splatting needs camera translation and parallax.
• Do not whip the camera around corners; slow down so frames remain sharp.
• Do not pause on one view for a long time; repeated frames add upload and processing cost without new coverage.
• Do revisit the starting area at the end when practical; a connected route helps alignment.

Step 1: upload the original video

Open Generate, choose Video, and upload an MP4, MOV, WEBM, or M4V file. There is no frame-extraction utility to install. Leighton analyses the clip and shows the projected number of sampled frames before generation.

Step 2: choose the Leighton model

• Leighton 1.0 Lite, $2 base and about 10 minutes; useful for testing a capture before spending more credits.
• Leighton 1.0, $4 base and about 20 minutes; the standard quality choice for most objects and spaces.
• Leighton 1.0 Max, $6 base and about 25 minutes; the larger training budget for demanding or dense captures.

Those base prices include the first 500 sampled frames. Above 500, every started block of 50 adds 10% of the selected model's base credits and projected time. The generation summary shows the final estimate before you spend anything. This linear adjustment pays for the extra alignment and GPU work rather than silently rejecting a large video.

Step 3: name, generate, and monitor

Give the splat a recognisable name, check the frame estimate and credit total, then generate. Upload and training status remain visible in the dashboard. Because the work runs on SplatMart's GPU, closing other programs on your laptop will not make training faster.

Step 4: clean up the PLY

When generation finishes, download the source PLY or open it directly in the free browser splat editor. Remove floaters, crop unused surroundings, check the scene from awkward angles, and export a clean delivery file. From there you can use it in a viewer or engine, publish it, or list it for sale.

When photos are better than video

Video is convenient, but photos give you deliberate control over shutter speed, exposure, and the exact viewpoints included. Use a photo set for professional camera rigs, drone mapping, dark interiors where motion blur is likely, or any scene where you need maximum sharpness. Leighton accepts up to 3,000 photos on the same Generate page. The photos-to-3D-model guide covers that workflow.

Why a splat may still fail

Cloud GPUs cannot invent viewpoints the camera never captured. A blurry video, an orbit with no high or low angles, changing light, or a room full of movement can still produce holes and floaters. First inspect the capture itself; retraining the same weak footage with a larger model rarely fixes missing coverage.

Frequently asked questions

How do I convert a video to a Gaussian splat?

Record a slow video with overlapping views around a static subject, upload the raw clip to a 3DGS generator such as SplatMart's Leighton, choose a model, and let the cloud service sample frames and train the scene. Download the resulting PLY and clean it in a splat editor.

Do I need to extract frames from the video first?

No. Leighton accepts MP4, MOV, WEBM, and M4V uploads and samples frames across the video automatically, up to 3,000 frames.

Do I need an NVIDIA GPU?

Not for cloud generation. Leighton runs training on SplatMart's server GPUs, so a normal Mac, Windows laptop, or other browser-capable computer can submit the job.

How many video frames can I use?

Leighton caps a video generation at 3,000 sampled frames. The first 500 inputs are included in the model's base price; larger working sets add credits and projected time in blocks of 50.

How long should my capture video be?

Long enough to cover every useful angle, but not padded with repeated views. A careful object orbit may take one to three minutes; a room or larger scene can take longer. Coverage and sharpness matter more than duration.

Can a video of moving people become a 3D splat?

A standard static Gaussian splat expects the scene to remain still. Moving people usually create duplicated or smeared artefacts. Capturing dynamic performance requires a 4D or time-varying reconstruction workflow, not this static pipeline.

Ready to use the clip already on your phone? Upload it to Leighton, review the frame and credit estimate, and generate the first version without installing a local training stack.