Turn 360° footage into immersive 3D Scenes

A plugin for LichtFeld Studio that turns 360° footage into a trained 3D Scene. It handles the whole chain — pulling frames out of your video, solving where the camera was, and preparing everything for training — so you don’t have to run separate tools for each step.

It’s free.

No. Everything runs on your own machine, on your own GPU. Your footage never leaves your computer.

Drop the plugin folder into LichtFeld Studio’s plugins directory and restart. That’s it — under 30 seconds.

No. That’s the point. Pick your footage, pick an output folder, press Generate.

The Insta360 X3 is what’s verified today — you can hand it the raw .insv files and it takes care of the rest. Beyond that, anything that produces a standard equirectangular 360° video will work, regardless of which camera made it. Other Insta360 models are likely to work but aren’t confirmed yet.

No. Moshpit360 reads the camera’s raw dual-lens files directly.

Either. An X3 clip is written as two files (_00_ and _10_, one per lens). Pick either one and the plugin finds its partner automatically — even if you’ve moved it to a nearby folder. You don’t need to select both.

Yes. Equirectangular video is supported. That said, the plugin is strongest when it works from the raw lens files, because it can skip stitching entirely.

Ordinary flat video or a folder of photos.

Yes. Projects save as a single small file and reopen with everything remembered, including what’s already finished — it won’t redo completed work.

Preparing frames takes minutes. Solving camera positions is the long part and scales sharply with clip length — anywhere from well under an hour to several hours for a long capture. Training then depends on your GPU and settings. It’s one click to start, not one click to finish.

Yes. Pause anywhere in the pipeline, including mid-training, and resume hours later from where you left off. You can also cancel outright.

Yes, but it will be using your GPU heavily. Avoid running other GPU-heavy work at the same time.

No — deliberately. It stops once reconstruction is finished so you can look at the result and choose your training settings before committing to a long run. Training is a separate, explicit step.

That’s expected behaviour, as above. Review the reconstruction, then start training when you’re ready.

A standard COLMAP dataset (your images plus the solved camera positions), and a trained Scene you can export as a .ply for SuperSplat, PlayCanvas, or any 3D viewer.

Almost always the capture rather than the software. 3DGS reconstructs a surface from parallax — the same surface seen from genuinely different positions. Walking once down a corridor gives you excellent rotational coverage but very little parallax for anything off that path, so those areas have nothing to reconstruct from. Fix it at capture time: walk through spaces rather than around them, cover each area from several positions, vary your height, and move slowly. Moshpit360 measures capture density and warns you before you spend hours on a run that can’t succeed.

Move slowly and steadily. Cross through the middle of rooms, don’t just skirt the edges. Cover important areas from more than one position and angle. Vary height between passes (waist and head level). Avoid fast turns and sudden motion.

Usually not. More frames along the same path add processing time without adding parallax. Covering the space from more positions helps far more than sampling the same walk more densely.

Yes. It picks a sensible number automatically so you’re not processing hundreds of near-duplicates, and you can override it whenever you want control.

A CUDA-capable NVIDIA GPU. It’s been developed and tested on an RTX 3050 with 8 GB, which works — more VRAM lets you train denser, more detailed splats.

It’s Windows-only today, since it runs inside LichtFeld Studio on Windows.

Lower the maximum Gaussian count in the training settings. The default target is aimed at larger cards; on an 8 GB GPU a lower cap will complete where the default won’t. There’s a trade-off — fewer Gaussians means less fine detail — so use the highest value that completes on your card.

That’s the warning described above. The reconstruction may still run, but expect a flat, shell-like result. Re-shooting with more coverage will help far more than any setting change.

There’s a feedback button in the plugin’s footer. It opens a short form with the technical details already filled in, so you only have to describe what happened.

Several GB per project. Intermediate frames are the bulk of it, and finished projects can be cleaned up afterwards.

Two halves of the same problem. UnrealTwin generates synthetic training data from Unreal Engine 5; Moshpit360 handles real-world captured footage. Building UnrealTwin taught us a lot about how 3DGS actually works, and Moshpit360 applies that to real capture.

It’s in beta and under active development. The full path from footage to a trained Scene works today; expect rough edges and expect things to change.

Join the beta. Testers get direct access to us and the in-plugin feedback button.