Three easy steps, one window, nothing in between: 1choose a folder of photos, a video or a lidar scan, 2press Run All, 3watch the scene take shape as it trains. LuluSplat extracts the frames, solves the cameras and trains a photorealistic 3D Gaussian Splat you can orbit, with no separate tools to install, entirely on your own GPU, on Mac, Windows or Linux. No uploads, no account, no subscription.
Windows may say “Windows protected your PC” for this installer — choose More info, then Run anyway.
The direct downloads are completely free.
Get it from the Mac App Store or the Microsoft Store to support development and receive automatic updates.
Mac · M1 or later · Metal · 179 MB
Windows · 10/11 · Vulkan · 247 MB
Linux (beta) · Vulkan · 262–265 MB
LuluScan: an even easier way to scan anything Coming soon
The scanner is the phone you already have. LuluScan is a companion app on the way for iPhone, iPad and Android: as you shoot, it records where the phone was for every photo, with LiDAR depth on devices that have it. The capture opens in LuluSplat with Import COLMAP Data and trains from the phone's own camera positions, with no camera solve.
See how a room is capturedHide
Fig. 01 — LuluScan · room guideFrom the app
You
Loop 1 · level
Coverage map
0% covered
Walk round the room with your back to the walls, pointing the phone across the room at the far side — not at the wall beside you. Go round once holding the phone level, then again tilting down at the floor and up at the ceiling. Step sideways between shots; don't turn on the spot.
Reading the map
While you capture, this map sits in the top corner of the camera. It shows what has been photographed so far; the percentage is how much of the capture is done.
A compass of where the camera has looked. Up is roughly the way you faced when the camera opened.
The middle band lights for shots held level, the outer band for shots tilted down at the floor, the inner band for shots tilted up at the ceiling.
The needle is the way the camera faces now.
Fill the middle band all the way round, then the floor. The ceiling counts least. The number of photos counts too.
Cameras from the phone Each photo keeps the position the phone's tracking gave it, so training starts without a camera solve.
LiDAR depth and normals On devices with LiDAR, depth and normal maps come with the capture and guide training.
Without LiDAR, too Where the phone's tracking is thin, LuluSplat fills in the point cloud from the photos and keeps the phone's cameras where they are.
Settings and up, included A focus region for an object, more iterations for a room, and the phone's gravity for which way is up.
LuluScan is in development. The animation is its guide for a room, as the app plays it before a capture: the walk on the left, the coverage map filling on the right. A video capture from LuluScan opens with New from Media and is solved like any other.
Fig. 02 — media in, splat and mesh out
A full run in LuluSplat 1.3.2 on a Mac mini M4 with 24 GB: 185 photos to a trained splat to a textured mesh.
01
Ingest
Stills, video, 360° photos and video, Insta360 .insv, or a lidar scan — mixed in one scene.
02
Extract
Frames pulled from clips, blurry ones rejected at a threshold you set.
03
Solve
COLMAP, Depmap or Speedmap places the cameras and builds a point cloud, checked before training starts. A lidar scan brings both with it.
04
Train
Vanilla or MCMC on your own GPU — Metal on Mac, Vulkan on Windows and Linux.
05
Export
.ply and .splat, an optional textured .glb mesh and .obj collider, and a run log.
01 How it works
Three steps from footage to a scene you can orbit.
Step 01 / 03
Drop in your media
A folder of stills, ordinary video, 360° video and photos, Insta360 dual-fisheye clips, or an E57 lidar scan with its 360° photos — several at once, mixed together, in a single scene. Camera positions are solved for you, or taken straight from the scan.
Shot on a phone, an action cam, a DSLR or a lidar scanner — it all works.
Step 02 / 03
Problems caught early
Blurry frames are dropped as they are extracted, and the reconstruction is checked before the long training step. A capture with too little overlap is flagged in seconds rather than an hour in.
A blur threshold you can tune, and a warning that says plainly what went wrong.
Step 03 / 03
Watch it converge
Orbit the scene as it trains — rough at first, sharpening as it goes. Switch the live viewport off for a faster run.
Full console output, and a RAM readout for the run in progress.
02 Features
Built for the whole capture, not just the pretty part.
01
Photos, video, 360° and lidar
Stills, ordinary video, 360° video and photos, Insta360 dual-fisheye clips, or an E57 lidar scan with its 360° photos — several at once, mixed, in a single scene. Both lenses are solved, and the camera operator is masked out automatically.
02
Problems caught early
Blurry frames are dropped as they are extracted, and the reconstruction is checked before the long training step — so a thin capture is flagged in seconds, not an hour in.
03
Two ways to train
Vanilla grows splats wherever detail is missing. MCMC trains a fixed number and keeps moving them to where they help most, with memory known before it starts. Train a project with each and compare.
04
Native on Mac, Windows and Linux
Reconstruction and rendering run on your own GPU — Metal on Apple silicon, Vulkan on Windows and Linux. No cloud instance, no queue, nothing to rent.
05
Three ways to solve the cameras
COLMAP, the proven choice; Depmap, which reads depth from every photo and copes with captures that barely move; or Speedmap, a fast on-site check of your frames. Already have COLMAP data? Import it and skip straight to training. A lidar scan skips the solve too: the scanner already knows where every photo was taken.
06
Completely private
Your photos, video and scans never leave the machine. Nothing is uploaded, there is no queue to wait in, and it all works offline.
Lidar scans: no camera solve
Open an E57 lidar scan with its 360° photos, such as an FJD Trion export, and LuluSplat trains from it directly. Each photo is placed where the scanner took it, so there are no cameras to solve. Training starts from the scan's own points, and its measured depth guides it. An FJD Trion P2 capture folder opens as it comes off the scanner, with no export step, and 360° or fisheye footage of the same walk is placed among the scanner's photos.
Fig. 03 — Lidar points
The scan · measured, not estimated
The points come from the scanner itself. Training starts from them instead of from a solved point cloud, so the columns, ceiling and floor are in place from the first step.
Fig. 04 — Splat
The splat · the result
The trained scene from the same viewpoint. Every photo sat where the scanner took it, and the scan's depth kept the walls, ceiling and floor where they were measured.
COLMAP and Depmap, side by side
A carved chamber at Medinet Habu, solved both ways. This capture is a hard case: there's no lidar scan, so the cameras have to be solved from the photos alone, and every photo was taken from the middle of the room, turning on the spot to face the walls.
Fig. 05 — COLMAP
117,702 points · 139 cameras
The red circle marks where the cameras are: all at the centre of the room. With almost no movement between shots, feature matching has little to triangulate from, and the walls come out sparse and broken up.
Fig. 06 — Depmap
625,061 points · 145 cameras
Depth is measured in each photo on its own, so it doesn't need the camera to move. Depmap places 145 cameras to COLMAP's 139, and the chamber comes out as solid walls, with over five times the points.
Depth guidance · normal guidance · finer growth · anti-aliasing
Result
scene.ply · 1.24 GB
Machine
Mac mini M4, 24 GB · LuluSplat 1.2.2
Depth and normals, read from every photo
For each photo, LuluSplat estimates how far away every pixel is and which way every surface faces, using the MoGe-2 model on your own GPU. MoGe-2 is built into LuluSplat, so there's no model to download and nothing extra to set up. Training can then lean on that geometry instead of guessing it from colour alone.
Fig. 08 — Depth
Depth · nearer is lighter
How far each pixel is from the camera. Depmap uses it to place the cameras and seed a dense point cloud, and depth guidance pulls the scene's shape toward it.
Fig. 09 — Normals
Normals · colour is direction
Which way each surface faces. Normal guidance turns every splat to lie flat along those surfaces instead of floating at odd angles.
Fig. 10 — Splat
The splat · the result
The trained scene: photorealistic, and ready to orbit, export as a splat, or turn into a mesh.
Fewer floaters Depth keeps splats on real surfaces, so there's less haze hanging in empty air.
Flatter walls, cleaner edges Walls, floors and roofs come out as surfaces, not lumpy clouds.
A head start Depmap begins training from a far denser point cloud, so the subject is covered from the first step.
Better meshes Splats that face the right way make a cleaner textured mesh and collider when you export one.
Depth guidance and normal guidance are quality options, off by default — switch them on per project.
Auto masking: the operator left out
In 360° and fisheye footage, whatever carries the camera is in every frame and moves with it: the person holding it, the stick, the helmet or the drone. LuluSplat finds it on its own, for each lens and each view of a 360° frame, from what stays put in the picture from frame to frame and where a person keeps appearing. It's left out of the camera solve and of training. People who appear only now and then, the ones being filmed, are kept.
Fig. 11 — Frames view · masks tinted
An Insta360 .insv walk through a kitchen, 148 frames. Tinted is what's left out: the edge of the lens, and the operator under the camera. In a few frames only the stick is caught. Outlined on the left: the Obstructions settings, with Find automatically on and the sliders for what it misses.
On by default Find automatically sits under Prepare frames ▸ Obstructions, and runs while the frames are made. Lidar scan projects mask the scanner and its operator the same way.
A steadier solve Features on the operator travel with the camera and pull the solve about. Masked, they're never matched.
Missed frames clean themselves up Where a frame slips through, the operator is in that photo alone. Every other photo of the spot shows the scene behind them, so training treats them as noise: their splats fade and are pruned.
See it at a glance The Frames view tints what's left out, so a mask that missed something shows at once. Masks turns the tint off.
For what the masks miss on fisheye footage, the sliders under Obstructions — Hide below, Crop bottom, Trim edges, Crop edges — leave out the bottom or the rim of every frame, from the solve or from the splat.
Speedmap: quickly verify your frames on site
A third solver, made for fast checks in the field. Right after shooting, Speedmap finds features on smaller images, uses about 300 of a long video's frames, and places every camera at once, then lists any frames it couldn't place. You know whether the capture is complete while you're still on location, and can reshoot the gaps before you leave.
Fig. 12 — left: Speedmap · right: COLMAP, global mapper
Speedmap
8 min 44 s
photos to trained splat · camera solve about 1 min
COLMAP · global mapper
15 min 33 s
photos to trained splat · camera solve about 7 min
The Speedmap run is on a Mac mini M4 with 24 GB. The same 185 photos and the same training on both: Vanilla, 7,000 iterations, a 1M splat budget. Speedmap's cameras are rougher, so its scene comes out softer — use it on site to confirm coverage, then solve with COLMAP or Depmap for the final scene.
03 Automation
Run it at volume, or without sitting in front of it.
Everything above is the whole app. These are for the days when one capture at a time is not the job.
A / Batch
Queue the captures and walk away
Line up as many captures as you like and leave the machine to it. Each becomes its own project with its own settings.
The list saves itself — a run interrupted overnight picks up where it stopped, finished projects skipped and failed ones retried.
Folders that already hold a COLMAP reconstruction go straight to training.
Ordinary video, 360°, and Insta360 clips can all go in together.
B / Command line
The app is its own CLI
Give it a subcommand and it runs headlessly instead of opening a window.
Build a capture, import a reconstruction, train again, or run a saved batch — from a script, a cron job, or a watch folder.
Progress on stderr, results on stdout: --json pipes straight into build systems and render farms.
Authorise a folder once and it works everywhere beneath it.
run
import
train
batch
grant
mcp
C / AI assistants
A built-in MCP server
Claude and other assistants can drive the pipeline themselves, over the Model Context Protocol.
Build a splat, follow it as it trains, and report on projects already finished.
Ask it how to shoot a capture that reconstructs well, before you shoot it.
Training logs are exposed as resources, so an assistant asked why a run was slow can read the timings rather than guess.
04 Export
Finish here. Keep working anywhere.
Save finished scenes as .ply and .splat — the formats Gaussian Splat viewers, web players, and DCC plugins already read. Turn on mesh export and training also saves a textured scene.glb for game engines and 3D apps, and a low-poly collider.obj for physics, both in the scene's own coordinates. Every run also appends its splat count, quality, duration, and peak memory to output-log.txt.
.ply
.splat
.glb
.obj
Splat viewers
Web players
DCC plugins
Need a different 3DGS format? Use our online converter,
Pebee
Not required — fully offline capable. No other tools to download.
What's new
1.3.2 latest
Occlusion-aware growth — Vanilla can count only the photos a splat is actually drawn in, not those where a wall or pillar hides it, so large indoor captures keep growing instead of stopping early. Under Quality options, off until you want it.
Paced growth — on a thousand photos or more, Vanilla paces its resets and growth by the number of photos, so it grows to its whole splat budget. With depth and normal guidance it's comparable to MCMC at higher iterations. No change under 300 photos.
See the frames — while Prepare frames runs, the viewport shows each frame as it's written, with what a mask leaves out tinted; afterwards a Frames button sits beside Sparse and Splat. Images the solve couldn't place are shown along the sparse view and marked in orange. Hover a frame and click the minus, or press E, to leave it out of the next solve and training — your files stay where they are.
Queue the next stage — while a stage runs, the stages after it offer Queue, and each starts once the one before finishes. Remove takes one out again, so a Run All can stop after the solve.
Larger budgets and captures fit — training's smaller photo copies are made a few at a time, and photos that don't fit beside the splats are read from disk as needed: a thousand-photo capture at 1600 px now trains at a 7M budget on a 24 GB Mac, where it stopped at 6.4M. The memory check counts depth and normal maps too.
A 7M splat budget, and advice for large captures — 7 million sits between 5M and 10M. From a thousand images up, the Trainer setting suggests MCMC on a long run, or Paced growth with depth and normal guidance to stay with Vanilla.
Large scenes show after training — when a run used nearly all of the GPU's memory, such as a 10M budget on a 24 GB Mac, LuluSplat now opens the saved scene itself instead of leaving the viewport empty.
1.3.1
LuluScan captures — captures from LuluScan, the upcoming iPhone, iPad and Android companion, open directly. Import COLMAP Data trains a photo capture from the phone's own camera positions with no camera solve, taking its LiDAR depth and normal maps and suggested settings with it; where tracking was thin, the point cloud is filled in from the photos.
Background — under Quality options. Random trains over a new colour each step, so dark walls, shadows and corners come out solid rather than see-through, for objects and rooms. Sky learns a background with the scene, for outdoor scenes that look out at the sky.
Sharper growth and spike cleanup — large splats are split over fine texture, and splats no longer stretch into spikes across new viewpoints. On for new projects: together they trained a test room as sharp as before with half the splats, in half the time.
Local tone correction and camera refinement — a per-photo correction that varies across the frame, for a phone's HDR and local tone mapping; and each camera may move slightly to where its photo says it was, for phone-tracked captures such as LuluScan imports.
Set which way is up — the Camera panel stands up a scene that came out on its side: Level to ground, a quarter turn at a time or upside down, or back to Auto. It's kept with the project, so nothing has to be reconstructed again.
Fixes — portrait videos come out upright; a lidar scan with a single panorama trains at its own size; E57 scans from scanners other than FJD's line their panoramas up with the scan; Depmap keeps a usable lens when it places only a few photos; and Sparse reconstruction shows how long COLMAP's final adjustment has run.
Clearer errors — cancelling no longer shows a "Pipeline error" dialog, and real errors say what failed, such as "Sparse reconstruction failed". Running Sparse reconstruction again on an imported COLMAP project no longer stops with "Only 0 image(s) available".
1.3.0
Lidar scans — open an E57 scan with its 360° photos, such as an FJD Trion export, and train with no camera solve: each photo is placed where the scan took it, training starts from the scan's points, and its measured depth guides it.
Add your own 360 footage — film the same walk on a 360 camera such as an Insta360 X5 and it's placed among the scanner's photos; by default only yours train.
Captures straight off an FJD scanner — open an FJD Trion P2 folder — its .las.tgz and Insta360 X5 video — with no FJD Trion Model export; the video is placed in the scan by the scanner's own trajectory.
What carries the camera, left out — in 360° and fisheye footage, the stick, drone or person holding the camera is found and left out of the solve and training. People being filmed are kept.
360° photos — flat 2:1 360° photos are split into views like 360° video, and Insta360 .insp photos are read lens by lens.
Fisheye lens edges — each fisheye frame trains as five views, up and down included, keeping nearly all of the lens; raise Iterations as the app suggests.
Fixes — zooming walks on into a room, panning close in no longer crawls, camera frustums stay readable on long walks, and lidar scans open level and train their floors.
1.2.3
One lens per camera — photos from several cameras in one folder are told apart by size; with Single camera on, each camera shares one lens, and portrait shots are turned to match.
Depmap solves photos with no lens information — each lens starts from the depth model's focal length — a 1,169-photo, five-camera capture now places 1,079 photos instead of 15.
Misplaced cameras are left out — photos placed far from the rest are dropped before training and named in the log.
Faster Depmap — the depth model runs once per photo, saving about 0.4–0.8 s a photo.
Speedmap: a quick check of coverage — a third solver for testing footage — about a minute on a 583-frame video, against eight for a full solve.
More accurate dense points — kept only where two views agree on their depth; the worst points are about half as far out.
Depth maps fitted across the picture — matched to the solve region by region; dense points now cover 79–99% of the scene instead of 59–79%.
Time left, and all the memory in use — per-step time left in the solve, depth and normal map progress, and a memory readout that includes COLMAP.
Fixes — no more shrunken sparse view, stale splat view, memory growth with depth or normal guidance, lost subjects beyond doorways, distortion stalls, or depth streaks.
1.2.2
Better rooms shot from one spot — photos taken from one place — turning inside a room, or from a tripod — now train cleanly; Depmap keeps far corners and what lies past a doorway, and plain walls and bright openings are no longer pruned away.
Fewer holes — dense starting points fill views they used to miss, such as a doorway to a bright exterior, and anti-aliasing no longer leaves black patches.
Solid meshes — textured meshes and colliders are fitted as one continuous surface, without holes in walls seen from one side, and their textures are free of stray specks.
MCMC uses all the starting points — it starts from an even sample of every point, Depmap's dense points included.
Every photo trains by default — Hold out test views is now off for new projects; turn it on when you want a quality score.
Camera controls in the viewport — a Camera panel sets the field of view for the sparse, splat and mesh views, and Reset view returns to the original framing.
1.2.1
Normal guidance — a new quality option turns every splat to match the surface direction in each photo, for better-oriented surfaces — made for the mesh exports. Off by default.
Better depth, from a new model — Depmap and depth guidance now use MoGe-2, which measures closer to the solved scene and supplies the surface normals.
Continue training from a checkpoint — every run leaves a checkpoint, so the next run picks up where the last left off — and can grow into a larger splat budget. About 700 MB per million splats; can be switched off.
Exposure and colour after training — exposure, temperature, tint, contrast and saturation for a finished scene, shown live and baked into the exports — or straight away with Write to scene files.
Focus on the subject — Focus region moves splats from the background to what your cameras circle — several times as many on the subject. Tight, medium or wide; works with Vanilla and MCMC.
Meshes without retraining — make a collider or textured mesh from a finished scene, and see it in the new Mesh view.
Fixes — Depmap no longer crashes on partly placed cameras; mixed cameras in one folder each get their own lens; impossible lenses no longer crash the solve; crash reports are kept on your Mac — nothing is sent.
1.2.0
Depmap: a second camera solver — estimates depth for every image, on your Mac, to place the cameras. It handles captures where the camera barely moves and adds up to 300,000 depth-measured points to start from.
A faster way to solve large captures — COLMAP's global mapper places every camera at once — a third less solve time on a 185-photo capture. The incremental mapper stays the default.
Four quality options for training — depth guidance for cleaner shape, finer growth for more detail, anti-aliasing for close-up views, and exposure compensation for footage that drifts. All off by default.
Distant subjects now train — a capture shot from a tight group of positions toward a faraway subject no longer trains to nothing.
A truer time left — the estimate accounts for the quality options and dense starting clouds, and is much closer early in a run.
Smaller fixes — solving again clears the old scene from the viewport; batches no longer draw live views; the "thin reconstruction" warning counts every starting point.
1.1.4
Textured mesh export — training can also save scene.glb, a textured 3D model for game engines and 3D apps, coloured from the splats, with a 2K or 4K texture. Best on solid things.
Mesh collider export — a low-poly collider.obj of floors, walls and objects for physics, in three levels of detail, in seconds.
Watch the cameras being solved — the point cloud builds up in the viewport as the cameras are solved, with the current step and its progress.
A level view — the point cloud and orbit view keep the horizon level, even when every photo looks down.
Vocab tree matching on the GPU — large unordered captures solve faster; CPU matching now shows progress, and very large sequential captures close their loops more reliably.
1.1.3
A new Vanilla trainer — runs on the same Metal engine as MCMC with gsplat's proven settings: dense detail trains in full, using less memory, and an Unlimited budget stops at what your Mac can hold.
The viewport fills the window — a wider window shows more of the scene instead of black bars.
A realistic time left — the estimate allows for training slowing down as the scene grows; a "~" marks it while it may still change.
1.1.2
Faster camera solving — feature matching now runs on the GPU and checks every feature, for slightly more detail. Switch it off in the sparse settings to match on the CPU.
Leaner bundle adjustment — the scene is refined less often as it grows, cutting that stage by about a third with the same accuracy.
1.1.1
Large captures caught before they load — checked up front, with the Training resolution that will fit.
Vanilla stops before memory runs out — halts at what your Mac can hold and keeps everything trained so far.
The live viewport steps aside — if the GPU can't spare memory, it switches off and training carries on.
1.1.0
A second trainer: MCMC — trains a fixed splat budget, so memory is known before it starts. Pick Vanilla or MCMC per project.
Splat budgets — cap either trainer from 100k to 100M; a budget this Mac can't hold is refused up front.
Training resolution — one setting caps the image size both trainers work from.
Help on every setting — a ? beside each one explains what it does.
A sharper live viewport — on from the start, no dropped splats in dense areas, smooth while training.
Half the memory for Vanilla — training images held at a quarter of their former size.
Finished scenes open faster — viewing no longer loads the training images.
1.0.4
Command line — drive the app from a script, with JSON results for pipelines.
AI assistants — an MCP server, so Claude and others can build a splat and follow it as it trains.
Drone propellers — Trim edges keeps fisheye rims out of tracking; Crop edges keeps them out of training.
Readable over bright scenes — the scene panel and sparse/splat switch no longer wash out against a sky.
Stopping actually stops — cancelling during a load takes effect at once, and reads as stopped rather than failed.
Large captures load leaner — peak memory during load is down by about a third.
1.0.3
Fisheye reconstructions — COLMAP data solved with a fisheye model is undistorted on import and trains like any other.
Multi-camera and rig captures — per-camera folders are found wherever they sit; COLMAP rigs import unchanged.
Folders read all the way down — images and video are picked up from subfolders at any depth.
Version in the window header — without opening Help to find it.
Several videos in one scene — mix as many clips as you like — ordinary video and 360° together.
Batch processing — queue captures, each its own project; an interrupted run resumes where it stopped.
Batches accept COLMAP data — a folder holding a reconstruction goes straight to training.
Training log — splat count, quality, duration and peak memory appended to output-log.txt.
Leaner memory — fixed a leak that grew GPU memory for the length of a run — peak down by about two thirds.
1.3.2 latest
Occlusion-aware growth — Vanilla can count only the photos a splat is actually drawn in, not those where a wall or pillar hides it, so large indoor captures keep growing. Under Quality options, off until you want it.
Paced growth — on a thousand photos or more, Vanilla paces its resets and growth by the number of photos, so it grows to its whole splat budget. With depth and normal guidance it comes close to MCMC. No change under 300 photos.
See the frames — while Prepare frames runs, the viewport shows each frame as it is written, with what a mask leaves out tinted. Afterwards Frames sits beside Sparse and Splat; click a frame to see it larger. Frames the solve could not place are shown along the sparse view and marked in orange.
Leave a frame out — hover a frame and click the minus, or press E in the larger view, to leave it out of the next solve and training. Your files stay where they are.
Queue the next stage — while a stage runs, the stages after it offer Queue, and each starts once the one before it finishes. Remove takes one out again, so a Run All can stop after the solve.
And more — a 7M splat budget, a suggestion of MCMC for large captures, less memory for photos while training, large scenes that open by themselves after training, and a tidier log: finished steps in green, paths shown from the project folder.
1.3.1
LuluScan captures — captures from LuluScan, the upcoming phone and tablet companion, open directly. Import COLMAP Data trains a photo capture from the phone's own camera positions, with its LiDAR depth and suggested settings; where tracking was thin, the point cloud is filled in from the photos.
New quality options — Background: Random trains over a different colour each step, so dark walls and corners come out solid; Sky learns a sky behind outdoor scenes. Sharper growth and Spike cleanup are on for new projects. Local tone correction handles a phone's HDR, and Camera refinement lets each camera settle where its photo says it was.
Set which way is up — if a scene comes out on its side, the Camera panel stands it up: Level to ground, quarter turns, or Auto. It's kept with the project, with nothing to reconstruct again.
Fixes — E57 scans from other scanners line their panoramas up with the scan; portrait videos come out upright; Depmap holds on to a lens that runs out of bounds; the camera solve shows how long a whole-model adjustment has run; and Intel graphics with older drivers train without errors.
1.3.0
Lidar scans — open an E57 scan with its 360° photos, such as an FJD Trion export, and train with no camera solve: each photo is placed where the scan took it, training starts from the scan's points, and its measured depth guides it.
Captures straight off an FJD scanner — open an FJD Trion P2 folder — its .las.tgz and Insta360 X5 video — with no FJD Trion Model export; the video is placed in the scan by the scanner's own trajectory.
What carries the camera, left out — in 360° and fisheye footage, the stick, drone or person holding the camera is found and left out of the solve and training. People being filmed are kept.
360° photos — flat 2:1 360° photos are split into views like 360° video, and Insta360 .insp photos are read lens by lens.
Fisheye lens edges — each fisheye frame trains as five views, up and down included, keeping nearly all of the lens; raise Iterations as the app suggests.
Fixes — zooming walks on into a room instead of stopping short, panning close in no longer crawls, and camera frustums stay readable on long walks.
1.2.3
One lens per camera — photos from several cameras in one folder are told apart by size; with Single camera on, each camera shares one lens, and portrait shots are turned to match.
Lenses for photos without EXIF — with Single camera off, Depmap starts each lens from its depth model's focal length.
Misplaced cameras left out — photos placed far from the rest are dropped before training and named in the log.
Faster Depmap — the depth model runs once per photo — about half a second saved a photo.
Speedmap: a quick check of coverage — a third solver for testing footage: small images, about 300 of a video's frames, and a list of the frames it could not place.
Better depth maps and dense points — matched to the solve region by region; on bicycle, dense points cover 75% of the scene instead of 54%.
Time left, and all the memory in use — per-step time left in the solve, depth-map progress before training, and a memory readout that includes COLMAP.
Fixes — no streaks out of the cameras, a subject beyond a doorway is kept, runaway lens distortion no longer stalls, and the sparse view no longer shrinks to a dot.
1.2.2
Better rooms shot from one spot — photos taken from one place — turning inside a room, or from a tripod — now train cleanly; Depmap keeps far corners and what lies past a doorway, and plain walls and bright openings are no longer pruned away.
Fewer holes — dense starting points fill views they used to miss, such as a doorway to a bright exterior, and anti-aliasing no longer leaves black patches.
Solid meshes — textured meshes and colliders are fitted as one continuous surface, without holes in walls seen from one side, and their textures are free of stray specks.
MCMC uses all the starting points — it starts from an even sample of every point, Depmap's dense points included.
Every photo trains by default — Hold out test views is now off for new projects; turn it on when you want a quality score.
Camera controls in the viewport — a Camera panel sets the field of view for the sparse, splat and mesh views, and Reset view returns to the original framing.
1.2.1
Better depth, and normal guidance — Depmap and depth guidance use MoGe-2 on your graphics card; its surface normals drive Normal guidance, which turns each splat to match the surfaces — made for the mesh exports.
Focus on the subject — Focus region moves splats from the background to what your cameras circle — several times as many on the subject. Tight, medium or wide.
Colour grading — exposure, temperature, tint, contrast and saturation for a finished scene, shown live and baked into the exports — or straight away with Write to scene files.
Clearer point clouds — points show as fixed-size dots that stay crisp at any zoom.
Meshes without retraining — export a collider or textured mesh from a finished scene, now with Ultra detail and 8K textures, and see it in the new Mesh view.
Fixes — mixed cameras in one folder each get their own lens; impossible lenses no longer crash the solve; crash reports are noted on the next launch.
1.2.0
Depmap: a second camera solver — estimates depth for every image with Depth Anything, on your PC, to place the cameras. It handles captures where the camera barely moves and adds up to 300,000 depth-measured points to start from.
Four quality options for training — depth guidance for cleaner shape, finer growth for more detail, anti-aliasing for close-up views, and exposure compensation for footage that drifts. All off by default.
Distant subjects now train — a capture shot from a tight group of positions toward a faraway subject no longer trains to nothing.
A truer time left — the estimate accounts for the quality options and dense starting clouds, and is much closer early in a run.
1.1.4
Textured mesh export — training can also save scene.glb, a textured 3D model for game engines and 3D apps, coloured from the splats, with a 2K or 4K texture. Best on solid things.
Mesh collider export — a low-poly collider.obj of floors, walls and objects for physics, in three levels of detail, in seconds. Both meshes share scene.ply's coordinates.
Watch the cameras being solved — the point cloud builds up in the viewport as the cameras are solved, with the current step and its progress.
A level view — the point cloud and orbit view keep the horizon level, even when every photo looks down.
Faster camera solving — vocab-tree matching and sequential loop closure run on the GPU too, and the scene is re-adjusted less often — about a third less time.
A realistic time left — the estimate allows for the scene growing; a "~" marks it while it may still change.
Projects open faster — a project opens on its point cloud; the trained splat loads when you choose Splat.
Fixes — Batch's Add Captures reads subfolders and takes a single video, like New from Media; batches keep the viewport off; vocab tree no longer asks for a file it already has.
1.1.3
Folders, all the way down — New from Media takes a folder and picks up images and video from subfolders at any depth; projects already built inside are skipped.
Mixed captures solve correctly — ordinary photos alongside fisheye or 360° footage keep their own camera model.
Starts alongside overlay tools — Vulkan overlays such as OBS, FPS monitors and GeForce Experience no longer make the app quit on launch.
1.1.2
First Windows release — LuluSplat on Windows 10 and 11 (64-bit), on any Vulkan-capable GPU from 2016 or later.
Trainers — Vanilla (gsplat) and MCMC (metal-gauss Vulkan port).
Faster camera solving — feature matching runs on the GPU.
External COLMAP — use your own CUDA build of COLMAP, set in Settings.
External FFmpeg — use your own FFmpeg, set in Settings.
1.3.2 beta
Occlusion-aware growth — Vanilla can count only the photos a splat is actually drawn in, not those where a wall or pillar hides it, so large indoor captures keep growing. Under Quality options, off until you want it.
Paced growth — on a thousand photos or more, Vanilla paces its resets and growth by the number of photos, so it grows to its whole splat budget. With depth and normal guidance it comes close to MCMC. No change under 300 photos.
See the frames — while Prepare frames runs, the viewport shows each frame as it is written, with what a mask leaves out tinted. Afterwards Frames sits beside Sparse and Splat; click a frame to see it larger. Frames the solve could not place are shown along the sparse view and marked in orange.
Leave a frame out — hover a frame and click the minus, or press E in the larger view, to leave it out of the next solve and training. Your files stay where they are.
Queue the next stage — while a stage runs, the stages after it offer Queue, and each starts once the one before it finishes. Remove takes one out again, so a Run All can stop after the solve.
And more — a 7M splat budget, a suggestion of MCMC for large captures, less memory for photos while training, large scenes that open by themselves after training, and a tidier log: finished steps in green, paths shown from the project folder.
Linux — closing a frame's larger view no longer freezes the app (AMD's Linux driver lost the GPU device), and its button says "Show in Folder" and opens the frame's folder in the file manager.
1.3.1
LuluScan captures — a capture from LuluScan, the upcoming phone and tablet companion, opens with Import COLMAP Data and trains from the phone's own camera positions with no camera solve, taking its LiDAR depth and normal maps and suggested settings with it. Without LiDAR, COLMAP adds points from the given cameras, which stay where they are.
Background — under Quality options: Black; Random, a new colour each step, so dark surfaces train solid rather than see-through, for objects and rooms; or Sky, which learns a background with the scene for open outdoor views. LuluSplat's viewer draws the sky, and other viewers ignore it.
Sharper growth and spike cleanup — large splats are split over fine texture, and splats no longer stretch into streaks across new viewpoints. Both are on for new projects.
Local tone correction and camera refinement — a per-photo correction that varies across the frame, for a phone's HDR and local tone mapping (it replaces exposure compensation); and each camera may turn and shift slightly to fit its photo, for phone-tracked captures.
Stand a scene up — the viewport's Up row stands up a scene that came out on its side: level it to the ground, turn it a quarter turn or upside down, or go back to Auto. Exports keep the reconstruction's own frame.
Progress and errors — Sparse reconstruction shows how many images the mapper has placed so far, and errors say what failed in their title.
1.3.0
Lidar scans — open an E57 scan with its 360° photos, such as an FJD Trion export, and train with no camera solve: each photo is placed where the scan took it, training starts from the scan's points, and its measured depth guides it.
Captures straight off an FJD scanner — open an FJD Trion P2 folder — its .las.tgz and Insta360 X5 video — with no FJD Trion Model export; the video is placed in the scan by the scanner's own trajectory.
What carries the camera, left out — in 360° and fisheye footage, the stick, drone or person holding the camera is found and left out of the solve and training. People being filmed are kept.
360° photos — flat 2:1 360° photos are split into views like 360° video, and Insta360 .insp photos are read lens by lens.
Fisheye lens edges — each fisheye frame trains as five views, up and down included, keeping nearly all of the lens; raise Iterations as the app suggests.
Fixes — zooming walks on into a room instead of stopping short, panning close in no longer crawls, and camera frustums stay readable on long walks.
Linux — finding what carries the camera runs on the GPU through Vulkan, with a CPU fallback; the file dialogs list lidar scans and .insp photos; time left after continuing from a checkpoint is no longer hundreds of hours. Video frames use FFmpeg 6.1, so 1.3.0's new colour handling (which needs 7.1) is left out on Linux.
1.2.3
One lens per camera — photos from several cameras in one folder are told apart by size; with Single camera on, each camera shares one lens, and portrait shots are turned to match.
Lenses for photos without EXIF — with Single camera off, Depmap starts each lens from its depth model's focal length.
Misplaced cameras left out — photos placed far from the rest are dropped before training and named in the log.
Faster Depmap — the depth model runs once per photo — about half a second saved a photo.
Speedmap: a quick check of coverage — a third solver for testing footage: small images, about 300 of a video's frames, and a list of the frames it could not place.
Better depth maps and dense points — matched to the solve region by region; on bicycle, dense points cover 75% of the scene instead of 54%.
Time left, and all the memory in use — per-step time left in the solve, depth-map progress before training, and a memory readout that includes COLMAP.
Fixes — no streaks out of the cameras, a subject beyond a doorway is kept, runaway lens distortion no longer stalls, and the sparse view no longer shrinks to a dot.
1.2.2
Better rooms shot from one spot — photos taken from one place — turning inside a room, or from a tripod — now train cleanly; Depmap keeps far corners and what lies past a doorway, and plain walls and bright openings are no longer pruned away.
Fewer holes — dense starting points fill views they used to miss, such as a doorway to a bright exterior, and anti-aliasing no longer leaves black patches.
Solid meshes — textured meshes and colliders are fitted as one continuous surface, without holes in walls seen from one side, and their textures are free of stray specks.
MCMC uses all the starting points — it starts from an even sample of every point, Depmap's dense points included.
Every photo trains by default — Hold out test views is now off for new projects; turn it on when you want a quality score.
Camera controls in the viewport — a Camera panel sets the field of view for the sparse, splat and mesh views, and Reset view returns to the original framing.
1.2.1
Better depth, and normal guidance — Depmap and depth guidance use MoGe-2 on your graphics card; its surface normals drive Normal guidance, which turns each splat to match the surfaces — made for the mesh exports.
Focus on the subject — Focus region moves splats from the background to what your cameras circle — several times as many on the subject. Tight, medium or wide.
Colour grading — exposure, temperature, tint, contrast and saturation for a finished scene, shown live and baked into the exports — or straight away with Write to scene files.
Clearer point clouds — points show as fixed-size dots that stay crisp at any zoom.
Meshes without retraining — export a collider or textured mesh from a finished scene, now with Ultra detail and 8K textures, and see it in the new Mesh view.
Fixes — mixed cameras in one folder each get their own lens; impossible lenses no longer crash the solve; crash reports are noted on the next launch.
1.2.0
Depmap: a second camera solver — estimates depth for every image with Depth Anything, on your PC, to place the cameras. It handles captures where the camera barely moves and adds up to 300,000 depth-measured points to start from.
Four quality options for training — depth guidance for cleaner shape, finer growth for more detail, anti-aliasing for close-up views, and exposure compensation for footage that drifts. All off by default.
Distant subjects now train — a capture shot from a tight group of positions toward a faraway subject no longer trains to nothing.
A truer time left — the estimate accounts for the quality options and dense starting clouds, and is much closer early in a run.
1.1.4
Textured mesh export — training can also save scene.glb, a textured 3D model for game engines and 3D apps, coloured from the splats, with a 2K or 4K texture. Best on solid things.
Mesh collider export — a low-poly collider.obj of floors, walls and objects for physics, in three levels of detail, in seconds. Both meshes share scene.ply's coordinates.
Watch the cameras being solved — the point cloud builds up in the viewport as the cameras are solved, with the current step and its progress.
A level view — the point cloud and orbit view keep the horizon level, even when every photo looks down.
Faster camera solving — vocab-tree matching and sequential loop closure run on the GPU too, and the scene is re-adjusted less often — about a third less time.
A realistic time left — the estimate allows for the scene growing; a "~" marks it while it may still change.
Projects open faster — a project opens on its point cloud; the trained splat loads when you choose Splat.
Fixes — Batch's Add Captures reads subfolders and takes a single video, like New from Media; batches keep the viewport off; vocab tree no longer asks for a file it already has.
1.1.3
First Linux release — x86-64, on any Vulkan 1.1 GPU: a .deb for Ubuntu 24.04, an .rpm for the RHEL 9 family, or a tarball that installs without root.