• Better rooms shot from one spot Photos taken from one place — turning around inside a room, or from a tripod — now train cleanly. Depmap keeps the room's far corners and what lies past a doorway, and training no longer prunes away plain walls and bright openings partway through. • Fewer holes Dense starting points now 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 now fitted as one continuous surface, so walls seen from one side no longer come out full of holes. Their textures are free of stray white and black specks. • MCMC uses all the starting points MCMC now starts from an even sample of every point, Depmap's dense points included, rather than only the first ones in the file. • 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 in the viewport's lower-left corner sets the field of view for the sparse, splat and mesh views, and Reset view brings the camera back to where the scene was first framed.

WHAT'S NEW (1.2.1) • Normal guidance A new quality option turns every splat to match the surface direction seen in each photo, for better-oriented surfaces — made for the textured mesh and collider exports. Off by default. • Better depth, from a new model Depth for Depmap and depth guidance now comes from MoGe-2, which measures closer to the solved scene and also gives the surface normals that normal guidance uses. • Continue training from a checkpoint Every run now leaves a checkpoint when it ends or is stopped. Turn on Continue from checkpoint and the next run picks up where the last one left off instead of starting over — for a run you stopped early, or a finished scene you want to train further. Raise the splat budget and it continues too, going back to where splats are added and growing into the new budget. A run stopped part way and continued scores the same as one never interrupted. Checkpoints take about 700 MB per million splats and can be switched off. • Exposure and colour after training A new Exposure slider brightens or darkens a finished scene by up to two stops, Temperature and Tint sliders set its white balance — warmer or cooler, toward magenta or green — and Contrast and Saturation sliders adjust its tones and colour, down to black and white. All are shown in the viewport as you drag. With exposure compensation on, you can also give the scene the exposure and colour of any one photo instead of the average of them all. Both are baked into the exported scene — by the next run, or straight away with Write to scene files, which updates the textured mesh too. • Focus on the subject Turn on Focus region and training moves splats from the background to what your cameras circle — several times as many on the subject in our tests. Choose how far the subject reaches: tight, medium, or wide. It works with both Vanilla and MCMC, and is off by default. With the Depmap solver, Focus dense points also measures far more starting points on the subject, so training begins with it well covered. • Meshes without retraining Make a mesh collider or a textured mesh from a finished scene straight from the Output section, at the detail you choose, without training again — and look at it in the viewport's new Mesh view, lined up with the point cloud and the splats. • Fixes Depmap no longer crashes on reconstructions where a camera could not place some of its points. Photos from more than one camera in one folder — told apart by their size — are now each solved with their own lens instead of some being left out, and a stray image of an odd size, such as a screenshot, is pointed out by name. A camera solve that gives a photo an impossible lens no longer crashes; that photo is left out, with a note. If LuluSplat ever closes unexpectedly, it now saves a crash report and its session log on your Mac, and tells you where on the next launch — Help ▸ Show Logs and Crash Reports opens the folder. Nothing is sent anywhere.

• Depmap: a second camera solver Choose Depmap in the Sparse settings and LuluSplat estimates depth for every image with Depth Anything, on your Mac, then uses it to place the cameras. It handles captures where the camera barely moves between shots, including the same viewpoint photographed several times, and adds up to 300,000 depth-measured points for training to start from. Depth maps are saved beside the images. • A faster way to solve large captures With the COLMAP solver you can now choose its global mapper, which places every camera at once rather than one at a time. On a 185-photo capture the solve step took a third less time with the same cameras, and the saving grows with the photo count. The incremental mapper stays the default. • Four quality options for training Depth guidance pulls the scene's shape toward the depth maps, for fewer floaters and flatter walls. Finer growth adds splats more readily and for longer, for more detail. Anti-aliasing removes speckle when viewing closer than the photos were taken, and is built into the exported scene. Exposure compensation corrects each photo's brightness and colour, for footage that drifts. All are off by default. • Distant subjects now train A capture shot from a tight group of positions toward a faraway subject could train to nothing. It now trains normally. • A truer time left The time-left estimate now accounts for the quality options and for dense starting clouds, and is much closer early in a run. • Smaller fixes Running the camera solve again clears the old scene from the viewport. Batches no longer draw live views. The "thin reconstruction" warning counts every point training starts from.

• Textured mesh export Turn on Export textured mesh and training also saves scene.glb: the scene as a textured 3D model for game engines and 3D apps, coloured from the splats themselves. Choose the level of detail and a 2K or 4K texture. It works best on solid things — floors, walls, furniture, vehicles. • Mesh collider export Turn on Export mesh collider and training also saves collider.obj: a low-poly mesh of the scene's floors, walls, and objects for physics in a game engine or 3D app. It is measured from what the cameras saw, comes in three levels of detail, and takes seconds. Both are off by default and share the exported scene's coordinates. • Watch the cameras being solved The point cloud now builds up in the viewport while the cameras are solved, and the viewport shows which step is running and how far along it is. • A level view The point cloud and the orbit view now keep the horizon level, even when every photo looks down at the subject. • Vocab tree matching on the GPU GPU matching now covers vocab tree as well as sequential and exhaustive, so large unordered captures solve faster. CPU matching now shows its progress too. Very large sequential captures also close their loops more reliably.

• A new Vanilla trainer (msplat is dropped) Vanilla now runs on the same Metal engine as MCMC, with the proven settings of the gsplat project. Dense detail trains in full, it uses less memory, and when the splat budget is Unlimited it stops adding splats at what your Mac can hold instead of running out. Existing projects open and train with it unchanged. • The viewport fills the window The live view now always fills the viewport's width. A wider window shows more of the scene to the sides instead of black bars, and the point cloud and the splats stay framed alike. • A realistic time left Training slows down as the scene grows and the images reach full size, and the time left now allows for it from the start, rather than promising a minute for a five-minute run. A "~" marks it while it may still change.

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) Bug Fix • Large captures are caught before they load A capture too big for your Mac's memory could make the app quit as training started. It is now checked before any images load, and if it won't fit you're told which Training resolution will. • Vanilla training stops before memory runs out Vanilla grows splats as far as the scene asks. It now stops at what your Mac can hold alongside the images, and keeps everything trained so far. • The live viewport steps aside If the GPU can't spare memory for the live viewport, it switches off with a note in the console and training carries on. (1.1.0) New Features • A second trainer: MCMC • Splat budgets • Training resolution • Help on every setting • The live viewport, on from the start • Half the memory for Vanilla training • Finished scenes open faster

(1.1.0) • A second trainer: MCMC Choose Vanilla or MCMC for any project. MCMC trains a fixed budget of splats, moving faded ones to where the scene needs them, so memory is known before it starts. Available from the command line and to AI assistants too. • Splat budgets Cap how many splats either trainer uses, from 100k to 100M. A budget this Mac can't hold is marked in the menu and refused before any images load, rather than grinding into swap. • Training resolution One setting caps the image size both trainers work from — most useful for imported reconstructions that link full-size photos. Native keeps the images as they are. • Help on every setting A ? beside each setting explains what it does and what each choice means. • The live viewport, on from the start It now renders every scene through one sharper renderer: dense areas no longer drop splats, empty space is black, and orbiting stays smooth while training runs. The coarse-to-fine schedule keeps running with the viewport on. • Half the memory for Vanilla training Training images are held at a quarter of their former size, roughly halving what a large capture needs. • Finished scenes open faster Viewing a saved scene no longer loads the training images — just the scene and its cameras.

WHAT'S NEW (1.0.4) • Command line Drive the app from a script: run a capture, import a reconstruction, or run a saved batch, with JSON results for pipelines. The App Store copy authorises a folder once, then works beneath it. • AI assistants An MCP server, so Claude and other assistants can build a splat, follow it as it trains, report on projects already finished, and advise on a capture before you shoot it. • Drone propellers Trim edges keeps the rim of each fisheye lens out of camera tracking; Crop edges takes it out of the images that train. Each masking control now says whether it changes the solve or the splat. • Readable over bright scenes The scene panel and the sparse/splat switch no longer wash out against a sky. • Stopping actually stops Cancelling during the load used to be ignored until the whole scene had reached the GPU. It now stops as soon as the load ends, and says what it is waiting for. A stopped run reads as stopped, not as a pipeline error. • Large captures load leaner A scene is no longer held several times over while it loads, cutting peak memory by about a third — so a big project is less likely to push the Mac into swap.

v1.0.3 • Fixed Colmap rig support. • Fixed New from Media recursive dir. v1.0.2 • Insta360 .insv Dual-fisheye clips import directly, both lenses solved. The camera operator can be masked out so they don't train in as floaters. • Several videos in one scene Mix as many clips as you like — ordinary video and 360° footage together. • Batch processing Queue captures and leave the Mac to it, each its own project with its own settings. The queue saves itself, so an interrupted run resumes where it stopped: finished projects skipped, failed ones retried. • Batches accept COLMAP data A folder that already holds a reconstruction goes straight to training. • Training log Each run appends splat count, quality, duration and peak memory to output-log.txt. • Leaner memory Fixed a leak that grew GPU memory for the length of a run — peak usage down by about two thirds. A RAM readout in the status bar shows what the current run is using.

v1.0.2 • Insta360 .insv Dual-fisheye clips import directly, both lenses solved. The camera operator can be masked out so they don't train in as floaters. • Several videos in one scene Mix as many clips as you like — ordinary video and 360° footage together. • Batch processing Queue captures and leave the Mac to it, each its own project with its own settings. The queue saves itself, so an interrupted run resumes where it stopped: finished projects skipped, failed ones retried. • Batches accept COLMAP data A folder that already holds a reconstruction goes straight to training. • Training log Each run appends splat count, quality, duration and peak memory to output-log.txt. • Leaner memory Fixed a leak that grew GPU memory for the length of a run — peak usage down by about two thirds. A RAM readout in the status bar shows what the current run is using.