PUBLIC PRODUCT GUIDE
From drone footage to a usable Gaussian Splat.
A practical guide to capturing, generating, aligning and downloading your 3D scene. No account is needed to read this documentation.
1. Capture the scene from several positions
Keep the same buildings or landscape features visible across neighboring views. Move around the subject and capture the sides you want to reconstruct. Aim for steady movement, sharp frames and consistent exposure. Repeated views from one hovering position do not provide the same depth information as views from different positions.
Record overlapping passes of the same area. Distant backgrounds can be retained; cropping is not required. Avoid relying on moving cars, people or wind-blown foliage as the main source of matching features.
2. Choose supported source files
Video upload accepts MP4, MOV, MKV, AVI and M4V. A stitched 360 film with a 2:1 aspect ratio is recognised automatically and can be marked as a 360 film in the generation form; every frame is then split into six camera views that cover the whole sphere. File extensions identify the container, not the codec. Processing must be able to decode the footage, and browser duration checks must be able to read the file. If a recording cannot be read, export a compatible copy and keep the original.
For image-based generation, upload JPG/PNG photographs or complete 2:1 equirectangular panoramas from different positions. Photos and panoramas can be selected together. Use at least 10 original images; the account shows the applicable file-count and size limits.
3. Upload and select your recordings
Create an account, verify your email and create a project. Upload the recordings for one area, then select the files you want in the generation. The portal checks your available upload capacity and applicable model limits. To resume an interrupted upload, select the same local file again.
You can combine multiple overlapping videos in one scene. Keep unrelated locations in separate projects so their recordings are not accidentally selected together.
4. Set image sampling and quality
For videos, “seconds per image” sets the spacing between extracted frames. A shorter interval uses more frames; a longer interval uses fewer. Each selected video must produce at least 10 sampled images, and the extraction pipeline permits at most 10,000 images in total. These are processing limits, not quality targets. Photos are used directly rather than sampled as video frames.
Resolution labels refer to the longest image edge: Quick is 1280 px, Detailed is 1920 px and Ultra is 2560 px. Available quality levels and training steps depend on the plan. See Pricing for the current plan differences. A higher preset cannot recover a surface that was never captured.
Optional sky masking excludes detected sky from training. It can reduce sky artifacts while leaving the choice to use it with you.
5. Generate and inspect the model
Submit the job and follow its status in the project. Processing runs on remote GPU workers. One job per customer may be queued or processing at a time. When the model is ready, open the 3D viewer and inspect it from the viewpoints that matter for your project.
Check facades, near trees and the edges of the captured area. Missing or unstable details may call for better source coverage rather than simply more training.
6. Align to a map or FBX
Open the alignment tool and select a map reference or upload an FBX. Place at least three corresponding point pairs, distributed across the scene. Calculate the alignment, inspect the overlap and save it. You can refine the scale, rotation and placement without training another model.
The reported point-fit error describes the selected correspondences. It is not an independent statement of survey accuracy. Keep your reference units and coordinate assumptions consistent.
7. Download a Gaussian Splat PLY
Download from the model card. If you have saved an alignment, use the adjusted export for the project placement. A Gaussian Splat PLY includes attributes beyond an ordinary point cloud; importing only its point positions does not reproduce the splat rendering.
In Unreal Engine, use a compatible Gaussian Splat renderer that imports PLY and supports your engine version. For example, the MLSLabsRenderer project documents a PLY-based Unreal workflow. This is an example of a third-party renderer, not a verified DroneSplatter integration. Check that renderer’s requirements, then verify scale and orientation with your own model.
8. Share and keep a local copy
Create a sharing link when others should be able to explore the model without an account. PLY downloads are available on Free too. Keep a local copy of results you want to preserve: models may eventually be deleted after an account stops using the service.