Case 01
No GPU worth training on
A laptop, a Mac, or an office machine with integrated graphics. Capture is fine on any of them; training is not. This is most people.
Training as a service
You shot the dataset. We'll run the GPUs.
Training a Gaussian splat scene needs a lot of VRAM for a long time, and most people who can capture a good dataset do not have a 48 GB card sitting under the desk. Send us the images. We solve the cameras, train the scene, and send back files you can use — or host it for you as an embed.
Who this is for
Three versions of the same problem.
You already know how to capture. What you are missing is a machine that can hold the scene in memory for a few hours.
Case 02
A card that runs out of memory
An 8 or 12 GB card handles a small object and then dies partway through a room. Training crashes at the densification step, or you cut the quality until the result is not worth having.
Case 03
The hardware, but not the hours
You could train it, but the machine is doing something else for two days and the client wants it Thursday. We queue it and it comes back trained.
What we accept
Send it in whatever state you have it.
Earlier in the pipeline is fine. If you have already run the camera solve, we will use yours and skip a step.
Input
Image sets
JPEG, PNG or RAW, in one folder. Around 200 frames for an object, 300 to 600 for a room. Keep them at full resolution — we downscale if we need to, you cannot upscale later.
Input
Video
MP4 or MOV straight off a phone, camera or drone. We extract and cull frames ourselves, dropping the blurred ones. Shoot slowly and keep the exposure locked.
Input
COLMAP output
An existing sparse reconstruction — cameras.bin, images.bin, points3D.bin and the source images. We train straight from your poses.
Input
Point clouds and prior scans
A .ply point cloud, lidar output or a previous scan can seed the training and stabilise geometry, as long as the source photographs come with it.
What it costs you in hardware
Roughly what you would need to do it yourself.
Figures are for training at full quality without cutting resolution. Turnaround is from the moment the upload finishes.
| Dataset | Typical frames | VRAM to train locally | Our turnaround |
|---|---|---|---|
| Single object, turntable | 150–250 | 12 GB | Within 24 hours |
| One room or interior | 300–600 | 16–24 GB | 24–48 hours |
| Multi-room or large interior | 800–1,500 | 24–48 GB | 2–3 days |
| Exterior, site or drone flight | 1,500–3,000+ | 48 GB and up | 3–5 days |
The process
Four stages, and you can stop after any of them.
Take the files and go, or leave it with us and get an embed link. Both are normal.
01
Upload and triage
You send the dataset by upload link, WeTransfer, S3, Drive or a posted drive for anything very large. We look at it the same day and tell you whether it will train, what it will cost, and anything you should reshoot before we start.
02
Solve the cameras
Frames are culled for blur and duplication, then matched to recover camera poses and a sparse point cloud. This is the stage that fails on bad data, which is why we do it first — if it will not solve, you find out before the bill exists.
03
Train
The scene trains to a set iteration count on our hardware, with densification and pruning tuned to the dataset rather than left at defaults. You get a short report with the splat count, iterations and final PSNR so you can compare runs.
04
Deliver, or host
Files come back over a download link that stays live for 30 days. If you would rather not deal with hosting a 90 MB scene yourself, we can clean it, add waypoints and give you an embed URL instead — the same product as a full capture, minus the site visit.
Output
What comes back
- Trained scene as .ply, full quality, no watermark
- Web-ready compressed .splat for viewers like Supersplat or Babylon
- Level-of-detail variants if you asked for mobile support
- The camera solve, so you can retrain or extend it yourself later
- A training report: frames used, iterations, splat count, PSNR
Terms
Ownership and privacy
- Your dataset and the trained scene are yours — we claim nothing
- We do not use your data to train anything else or show it as a sample
- Source files are deleted 30 days after delivery unless you ask us to keep them
- Happy to sign an NDA before you send anything
- Nothing is published anywhere unless you ask us to host it
Before you upload
Why datasets fail to train.
Almost every failure is decided at capture time, not at training time. No amount of GPU fixes these.
Failure 01
Motion blur
Walking pace with a slow shutter produces frames that look fine on a phone screen and cannot be matched to each other. Shoot at 1/250 or faster, or move slower.
Failure 02
Auto exposure and white balance
If the camera re-meters every time you turn towards a window, the same wall arrives in three different colours and the scene trains muddy. Lock both before you start.
Failure 03
Not enough overlap
Aim for 70 % overlap between neighbouring frames. Sparse orbits and single passes down a corridor leave the solver with nothing to tie frames together.
Failure 04
Featureless surfaces
Blank white walls, plain floors, mirrored lobbies and glass corridors give the matcher no texture to work with. Capture more of the surrounding detail so it has anchors.
Failure 05
Anything that moved
People, pets, traffic, curtains and daylight shifting over an hour all become ghosts. If something had to move, tell us where and we will mask it out.
Failure 06
Missing angles
The solve can succeed while a corner nobody photographed still trains into cloud. Shoot the awkward heights: low over furniture, high towards the ceiling.
We check before we charge
Stage two is the honest gate. If the camera solve fails or the coverage is too thin to give a result worth paying for, we tell you what went wrong and what to reshoot, and there is nothing to pay. We would rather have the reshoot than your money for a bad scene.
Pricing
Per dataset, quoted after triage.
Price depends on frame count, resolution and whether you want the web-optimised variants. Repeat customers and batches of scenes are priced per run rather than per job.
Sending it
How to get the files to us
- Under 5 GB — WeTransfer, Drive or Dropbox link
- Over 5 GB — we send a direct upload link
- Very large or rural connection — post a drive, we return it
- Tell us the camera, the frame count and roughly what the space is
Next step
Send the dataset. Get the scene back.
Tell us the frame count, the camera and what the space is, and we will confirm whether it will train and what it will cost before anything starts.