If you shoot with an Insta360, the best file you can give us is the one straight from the camera: the raw .insv. We rebuilt our Gaussian Splatting pipeline around it. It now works on the original fisheye images, reads the camera's motion sensors, and returns scenes that are level and at real-world scale — faster and more precisely than from an exported video.
Every scene now also comes with a textured mesh optimised for floors, ready to drop into robot-training simulators such as MuJoCo and NVIDIA Isaac Sim.
Here is what changed, why it matters, and what to upload from now on.
What is new in the .insv pipeline
| Feature | What it means for your splat |
|---|---|
| Native fisheye processing | We work on the two original lens images, not on a stitched panorama. No stitching, no stitching errors. |
| Optimised loop closure validation | When your path comes back to a place it already saw, the pipeline checks the match more strictly. Long walks and multi-room captures line up better, with less drift. |
| Automatic metric scale | The scene comes out at real-world scale. One metre in the splat is one metre in the room — useful for measuring, for floor plans and for placing objects in AR/VR. |
| Levelled scene from IMU data | The .insv carries the camera's gyroscope and accelerometer data. We use it to find gravity, so the floor is flat and walls are vertical from the start — no manual rotation in the viewer. |
| Much faster processing | Skipping the stitched panorama removes a whole conversion step and gives the reconstruction cleaner input, so your scene is ready in a fraction of the time. |
Why the raw fisheye is better than an equirectangular video
An equirectangular video is what you get when you export from Insta360 Studio: the two fisheye images are stitched into one panorama and stretched onto a flat 2:1 frame. It looks fine to the eye, but the reconstruction sees the problems:
- Stitching aberrations. Along the seam between the two lenses, the panorama is an estimate. Lines bend, details double or disappear, and the error changes from frame to frame. Gaussian Splatting needs the same point to look the same from every viewpoint, so these errors become blur, floaters and warped surfaces.
- Stretched pixels. Near the top and bottom of the panorama (ceiling and floor), pixels are heavily stretched. The data looks bigger than it is, but no detail is added.
- No sensor data. An exported video usually loses the motion data the camera recorded. Without it, we cannot level the scene or recover its scale automatically.
With the raw .insv, each lens image is used as the camera saw it, with its real optics. That is why the results are more precise — and the processing is faster.
The equirectangular pipeline is still here
If you only have an equirectangular .mp4 — from another 360 camera, or because the original files are gone — you can still upload it. That pipeline has not changed: it works as before, and it produces good scenes from good footage.
We just do not recommend it when the .insv is available. Stitching always adds some error, and a stitched video gives none of the benefits above: no IMU levelling, no automatic scale, slower processing.
| Raw .insv (recommended) | Equirectangular .mp4 | |
|---|---|---|
| Input | Original fisheye images | Stitched 2:1 panorama |
| Stitching artefacts | None | Present along the seam |
| Level scene (IMU) | Automatic | No |
| Real-world scale | Automatic | No |
| Processing speed | Much faster | Standard |
| Supported cameras | Insta360 | Any 360 camera |
New: sim-ready meshes for robot training
Next to the splat, every processed scene now includes a textured mesh. It is built for one job in particular: putting a robot in it.
- Optimised floors. The walkable surface is the part a robot touches all the time, so we reconstruct it as a clean, continuous plane instead of a noisy, bumpy surface. Wheels and feet stay on the ground, and contacts in the physics engine stay stable.
- Plug and play. The mesh comes in standard formats that load straight into MuJoCo and NVIDIA Isaac Sim, so you can go from a walk-through video to a training environment without cleaning up geometry by hand.
- Real places, not hand-built rooms. Train and test navigation in the actual office, warehouse or home where the robot will work, with its real layout and clutter.
You will find three files in the output .zip:
| File | What it is |
|---|---|
model.glb |
The textured mesh, for 3D tools, the web and simulators |
model.usdz |
The same mesh in USD, for Isaac Sim and iOS (AR Quick Look) |
mesh_textured_0.png |
The texture atlas used by both |
Meshes are generated for every dataset, whatever the source. From an .insv they get the extra benefits of the new pipeline: the mesh is level and at real-world scale out of the box, so the floor is where the simulator expects it and distances match the real room. From other sources, you may need to scale and level the mesh yourself before training.
How to upload your .insv
- Copy the files from the SD card. Take them from the camera's memory card or the camera itself, not from an Insta360 Studio export. Each 360 clip is a single .insv file. Skip the
.lrvfiles: they are low-resolution previews. - Do not convert or re-export. Upload the .insv exactly as it came out of the camera. Any conversion throws away the fisheye images and the sensor data.
- Pick your camera on the upload page. Choose Insta360 as the source. The accepted format (.insv) is selected for you.
- Have an equirectangular .mp4 instead? Choose Equirectangular video. Uploads are sorted by file now, so an .mp4 always goes to the equirectangular pipeline.
The capture rules have not changed: walk slowly, overlap your path and keep exposure stable. Our Insta360 capture guide covers them in detail.
FAQ
Which cameras does the native fisheye pipeline support? Insta360 cameras that record .insv files, such as the X5.
I already exported my clip to .mp4. Do I need to shoot again? No. Upload the .mp4 as an equirectangular video and it will be processed as before. If you still have the original .insv on the card, upload that instead for a better result.
Does the .insv include both lenses? Yes. A single .insv clip contains the images from both lenses plus the motion data. Upload just that file.
Is the new pipeline available on the free plan? Yes. It is used automatically for every .insv upload, free or paid.
Do I get the mesh on the free plan too?
Yes. Every processed scene, free or paid, includes model.glb, model.usdz and the texture atlas in the output .zip.
Can I use the mesh in other engines?
Yes. .glb and .usdz are standard formats: the mesh opens in Blender and most 3D tools, and in Unity or Unreal Engine through their glTF importers.
Get a sharper, level, real-scale splat — and a mesh to train on
The short version: upload the .insv, not the export. You get native fisheye processing, a scene that is level and at real-world scale, a faster result, and a floor-optimised mesh you can load straight into MuJoCo or Isaac Sim.
Upload your .insv now → (Free tier, no install, no credit card.)