Release Notes

What changed in each version of Methodox Splats, newest first. The same account is in the program, under Help > What's new.

0.4.3 - 17 November 2026

The sections of the left panel fold away. Click a section's title, Capture, Cameras, Train, Objects or Deliver, to fold it, and again to open it. The program remembers which you folded, so a panel set up the way you work stays that way.

Help leads somewhere. Help > User guide opens the guide on the Methodox Wiki, Help > Sample scenes opens the scenes to download and look round, and Help > Website opens splats.methodox.app. Help > About is now just About, and has the same two links.

The licence notices listed in About are shorter: the one for video now covers only what this program contains.

0.4.2 - 17 November 2026

Reflections that touch the ground. A polished model mirrored the scene as if all of it were far away, so a chrome ball standing on a pavement showed the pavement floating inside it. A reflection is now looked for where it really lands: the joint in the paving runs to the point where the ball touches it, as it would on a real one, and what is near the model is the right size in it. Where the capture never looked, the model mirrors a soft blur of what is round that and no longer a flat grey patch.

Set the sun by its shadow. Under a sun, the selected model has a dot on the ground where its shadow ends. Drag the dot until the shadow runs the way the scene's own shadows do, and the sun is where it has to be, round and up at once. The two sliders are still there.

splats place --relight reads the scene's light again for the models a project already has, which is how a project lit by 0.4.1 gets the new reflections.

A project keeps a model whose file it cannot find. A project opened where one of its models' files is not, after being moved to another folder or another machine, used to forget that model the next time it was saved, with where it stood and how it was lit. It is now kept: the Objects stage names it with Locate… beside it to say where the file is, and it stands where it stood. Putting the file back where it was does the same.

What to know. At the edge of one thing in front of another a reflection can bend where it should break. A capture that saw only part of a room gives a reflection that is partly blur, and nothing here can put there what was never photographed.

0.4.1 - 17 November 2026

The scene lights a model as it is, and a sun can be set. A model is now drawn with the materials its file gives it. A part that is metal or smooth reflects the scene as it looks from where the model stands: a chrome ball put down in a street shows the street. And the Objects stage has a sun, because a capture remembers colours and not brightness, and its own sun is a white patch no brighter than its sky: set how strong it is and where, and the models throw shadows across the scene away from it and are lit from its side. Find looks for the sun in the scene's own shadows. On the command line, splats place --sun takes it by its angles, by the scene's shadows or by the brightest of the sky, and splats light writes what a scene looks like from a point as one picture, which is the light a model there is lit by.

It was looked at in three captures. In a street a chrome ball mirrors the street and a brass one shows it blurred. On a hill filmed from the air under a real sun, a ball and a helmet are lit from the side the ruined tower is and throw shadows the way the tower's falls. On a kitchen counter a ball is half hidden by the bowl in front of it.

Shadows that fall on what you see. The shade a model casts is now worked out for each pixel of the picture and not for each splat of the scene. On a capture filmed from the air the first way threw no shadow at all, the grass there being painted partly by splats far under the hill. The soft shade at a model's foot is sharper for the same change.

What to know. Find is a starting point: on the captures it was tried on it was not sure of any direction, and said so. A reflection is of the scene from the model's middle, so a large polished thing, or one close against a wall, mirrors the scene as if it were all far off. Nothing is drawn clear. The scene shades a model as a whole and not part by part.

0.4.0 - 17 November 2026

A 3D model in the scene. A .glb, .gltf or .obj can be stood in a splat and is drawn as part of it. In the studio, Add model in the new Objects stage or the File menu, or the file dropped on the view, stands it where the view is pointed; drag it to where it belongs, and set its size and turn on the bar under the view. Whatever of the scene is in front of it hides it, in a splat that moves as well as a still one, so a person walking past passes in front; it is lit by the scene where it stands; and it shades the ground at its foot. The models are saved in the project and drawn in an exported video. From the command line, splats place puts a model into a project by a camera and a pixel or by coordinates, and splats render draws the project with it.

It was stood in two captures and looked at. On the truck, a duck, a lantern and a radio stand on the paving from every camera they were looked at from, the truck's tail in front of the duck where it is in front; on the basketball scene a lamp post on the floor is hidden by the player and the ball as they pass behind it and stays put through all 150 frames.

What it does not do yet. A model's surface is drawn matt: nothing shiny or clear. Its shade is the soft one of an overcast sky, a tenth or so at its foot, and not a shadow cast by a sun, so under hard sunlight it looks less planted than it should. It does not move. Where a capture saw the ground poorly, a model stood there can sink a little into it. The viewport draws about a third slower while a model is in view.

0.3.2 - 16 November 2026

What happens between two frames. A splat that moves now moves between the frames it was filmed at. Each moving thing's splats are given the speed their thing is going, worked out before training from where it is in the frames before and after, so between two frames a thrown ball is one ball, part of the way along. In 0.3.1 it was two faint ones, which is why such a splat was played a frame at a time. Time now runs smoothly in the studio at whatever rate the window draws, a video can be exported slower than the scene was filmed, and splats render --frame 62.5 draws the half-frame. A project made with 0.3.1 plays a frame at a time as it did; train it again for the rest.

How true that in-between motion is was measured the one way it can be: train from every second frame and score the frames between, which the splat never saw, against their photographs. On the basketball scene those frames score 24.7 dB where something moves, against 25.9 at the frames it was trained on; 0.3.1's splats scored 20.4 there. On a second scene of that set, a man juggling, 24.0 against 21.2. splats animate --train-every 2 runs that test on any capture, and splats score reports it for a project trained that way.

Better at the frames themselves, and smaller if wanted. Because a splat is now in the right place at the frames either side of its own, it learns from their pictures too: the 150 frames of the basketball scene score 30.52 dB, and 26.15 where something moves, where 0.3.1 scored 28.94 and 25.33. --seed-every 2 starts splats at every second frame and lets their speed carry them between, for a project of 26 MB where it is otherwise 42, trained in two thirds of the time, for a tenth or two of a decibel where something moves.

Where little moves and it moves slowly, as in the kitchen scene, there was nothing to gain, and about a quarter of a decibel is lost (33.53 dB where 0.3.1 scored 33.99 and the same done without speeds now scores 33.77). A speed is given only where the frames before and after agree on it, which is what keeps that loss small; --no-flow trains without speeds, as 0.3.1 did.

0.3.1 - 16 November 2026

Things that move fast. A splat that moves is trained a new way, and what came apart before now holds together: a thrown ball, a hand, a person who crosses the room. The part of the scene that does not move is trained once, from each camera's picture of it with the moving things taken out, and each moving thing is given splats of its own for each instant, which appear and fade as time passes. Nothing is followed from one frame to the next, so nothing is lost for moving too far in a frame. On the basketball scene that 0.3.0 was shown on, all 150 frames, the four cameras kept out of training now score 28.94 dB where 0.3.0 scored 24.47, which is level with the figure published for that set, and the project is 43 MB where it was 147. The old way is still there as splats animate --method tracked, for when one splat has to be followed through a clip.

Training one in the studio. A scene filmed by several cameras at once now opens in the studio like any other capture: open its folder, and Train does the rest. The Train stage says which of four stages it is in, and while the motion trains the view runs through the clip, so what is forming is seen moving.

A cleaner room. Two things that are particular to a rig of fixed cameras are now handled, and between them they are worth three to four decibels on that scene, to the old way of training as well as the new. The cameras see each other, and what was drawing each camera's own housing sat in front of that camera's lens as a dark blot; nothing so near a camera is drawn from it any more. And each camera gets a colour correction of its own, since a rig's cameras are seldom matched (--no-exposure turns it off).

Scoring a moving splat. splats score project.msplat --capture <folder> scores every frame from the cameras kept out of training, over the whole picture and over only the part of it that moves. splats animate reports both.

One moment as a picture or a file. splats render project.msplat --frame 60 --out frame.png draws one moment, from the project's own cameras with --view <n>. Export in the studio's Deliver stage writes the moment showing.

It still needs a rig of cameras, not one camera moving. A thing that stays in one place for most of the clip is taken for part of the room. And such a splat is played at the frames it was filmed at: slowed down, it shows each frame for longer, not what happened between two of them.

0.3.0 - 15 November 2026

Splats that move. A scene filmed by several cameras at once can be trained into a splat that moves: splats animate <capture> --out <folder> trains the first instant as a still splat and then moves the same splats, frame by frame, to where the cameras next see them. It reads a folder of a rig's videos with a poses_bounds.npy, which is how the Neural 3D Video set comes, and a PanopticSports scene. The studio opens the project it writes and plays it: a bar under the view has play, pause (Space) and a slider for time, the camera is free while time runs, and File > Export video films the splat moving. splats render does the same from the command line, and --frame-files ply writes every frame as an ordinary splat file for other programs.

This is a first version, and it is plain about its limits. It follows slow motion well and smears what moves far between two frames, a thrown ball or a fast hand; it needs a rig of cameras, not one camera moving; and the studio plays a moving splat but does not yet train one. Help > Concepts has a section on how it works.

0.2.1 - 14 November 2026

A camera track for Blender, with no training. Once the cameras are solved, To Blender in the Cameras stage (also File > Export cameras to Blender, and splats track on the command line) writes a Blender scene: one camera that moves through the photographs in the order they were shot, with the photographs behind it as a film; a camera for each photograph with that photograph behind it; and the sparse points. Everything hangs from one object named Capture, which is what to turn, move and scale to set the scene on a floor at a size. It opens in Blender 4.2 and later. splats track --frames all keeps every frame of a video, for a track of the whole shot.

Getting to the cameras. The switch at the top of the view reads Orbit, Fly, Camera whenever there are cameras. Left and Right take any view onto the nearest camera and then to the one shot before or after, and Home and End go to the first and the last. The status bar says which camera the view is at.

The side panel is a tone darker than the stages on it, so each stage reads as a block.

What's new, this page, under Help.

0.2.0 - 13 November 2026

Help. Help > Concepts (F1) explains what Gaussian splatting is and the words the program uses. Help > About says the version, the GPU in use and the licences the program carries. The version is in the window's title.

The capture's cameras. Each camera is drawn where it stood, with the path through them in shooting order, over a finished splat as well as over the sparse points. Look through puts the view exactly where a photograph was taken, with a slider that fades the photograph in over the splat and Space to flip between the two.

Moving about. Fly (Shift+F): W A S D to move, E and Q to climb and sink, a drag to turn the head, Shift faster and Ctrl slower. The world's axes are drawn in a corner of the view. F11 shows the splat and nothing else on the whole display.

Renderer settings, from the button at the top right of the view: colour detail, antialiasing, background, and how a heavy view is drawn while it moves.

Training. Smoothing, on unless turned off, keeps splats from becoming finer than the photographs could show, which is what broke a result into needles up close. Exposure is for footage whose brightness or colour changes from frame to frame. Refine cameras adjusts the solved cameras as training goes.

Scoring. splats score says how closely any splat, from this program or another, reproduces a capture's photographs.

Fixed. A Blender file exported from Methodox Splats crashed Blender as soon as a viewport drew it; it now opens and draws. The window is redrawn while its edge is dragged. A project trained at the photographs' full size reopens saying so.

Changed. The stages are named Capture, Cameras, Train and Deliver, without numbers.

0.1.0 - 13 November 2026

The first version. A video, a folder of photographs or a capture solved elsewhere goes in; the program solves where each photograph was taken from, trains a splat on whatever GPU the machine has while you watch it form, and writes it as PLY, SPZ, SOG, glTF binary, compressed PLY, .splat or a Blender file. A project keeps the cameras, the splat, the settings and the view in one file. The view turns round the subject or follows a path of saved views, and either is exported as a video. H.264, HEVC, AV1, VP9, MPEG and ProRes video is read with nothing else installed. The splats command line does each step by itself. A Windows installer.