SplatsThe evolution of media, in brief
RSS
🔬Research🏢Industry🧰Tools📷Hardware🎨Culture

1 big thing

Two blurry photos are now enough for a 3D scene

CasDeblurGS reconstructs a coherent gaussian scene from exactly two motion-blurred images — no camera poses, no sharp reference frames, no per-scene optimisation.

Daniel Habib

Daniel Habib

2 hours ago-Research

Two smeared input photographs of a black car, beside the method's reconstructed view and the ground truth, with the grille crop enlarged underneath each
A query photo of a red handbag among clutter, beside a baseline extraction where the bag comes out surrounded by smeared background, beside Seed2GS's clean isolated bag on black

Pull an object out of a splat you didn't capture

3 hours ago

Two aerial views of the same reconstructed conifer forest: at ten simulated minutes a small orange and black burn scar, and at fifty minutes a black scar covering most of the frame

They set fire to a gaussian splat of a real forest

5 hours ago

Three columns: the raw compressive measurement as an unreadable scatter of speckle, the method's reconstruction of a plate of hot dogs and a vending machine, and the ground truth beside it

One exposure in, a whole 3D scene out

Yesterday

The same orchard branch twice: on the left the radiance field rendered photorealistically, on the right the semantic field, with apples picked out in red and foliage in green

Radiance fields stop being pictures and start being places

Yesterday

Daniel Habib

Daniel Habib

Yesterday-Research

Volumetric video turns privacy into a 3D problem

Three living-room scenes in a grid: the original capture, the ground-truth mask with one person filled in red, and the system's predicted mask filling the same person in blue

Hossein Khalili and colleagues at UCLA and Nokia Bell Labs have built InViStream, a system that removes private content from RGB-D streams on the camera side, before anything is fused into a shared 3D scene.

Why it matters: Ordinary video can be redacted frame by frame. A volumetric pipeline can't be, because the thing you are trying to hide is reconstructed from several cameras at once — miss it in one view and it reappears in the fused scene.

Go deeper (2 min. read) ⟶
Daniel Habib

Daniel Habib

2 days ago-Research

LocusGS gives instant splats a sense of place

Two reconstructions of the same scene. In the TokenGS result the orange gaussians decoded from one token are scattered across distant parts of the scene; in the LocusGS result the same orange points sit in one tight cluster

Wenyu Li and colleagues at NUDT diagnosed a structural flaw in query-based feed-forward gaussian splatting — the queries have no idea where they are — and fixed it by giving each one an explicit position in 3D space.

Why it matters: Feed-forward splatting is the branch that skips per-scene optimization: images in, gaussians out, no training run per capture. It is the path to splats that appear as fast as you can photograph something.

Go deeper (1 min. read) ⟶
Daniel Habib

Daniel Habib

2 days ago-Research

Splatting leaves graphics and turns up in the CT scanner

A walnut with metal pins inserted, beside two CT reconstructions of it: the conventional FDK result streaked with bright artifacts, and the paper's result showing the same slices clean

Kiseok Choi, Min H. Kim and colleagues have adapted gaussian splatting into a reconstruction method for cone-beam CT, correcting the streaks and dark bands that metal implants throw across a scan by modeling the polychromatic X-ray beam that causes them.

Why it matters: Splatting was built to render what a camera saw. This is the same representation running in reverse, as a solver for a physical inverse problem in a medium that has nothing to do with photography.

Go deeper (1 min. read) ⟶
Amara Osei

Amara Osei

2 days ago-Hardware

Gaussian splatting is getting its own silicon

Two rows of splat renderings of a locomotive: the top row at fp32 precision, the bottom row at int8 without cross-correlation, where the scene softens and detail smears

A team led by Siddarth Gottumukkula and Priyesh Shukla has proposed ProbSplat, a compute-in-memory chip architecture that stores gaussians as physical charge and evaluates their log-likelihoods in analog — targeting the power budget of standalone AR/VR headsets and edge robots rather than a desktop GPU.

Why it matters: Every splat pipeline shipping today assumes a GPU somewhere — on the device, or in the cloud with a round trip. That assumption is what keeps live reconstruction off untethered headsets and small robots.

Go deeper (1 min. read) ⟶
Theo Lindqvist

Theo Lindqvist

3 days ago-Tools

Splatlight brings self-shadowing splats to Blender

Split view of a toy robot and plush scene shown as raw gaussians on one side and relit in Blender by Splatlight on the other

Splatlight, a new Blender add-on from developer Javier Herrera, makes imported gaussian splats respond to scene lights and cast self-shadows — ending the era of splats that carry their capture lighting baked in and ignore every lamp around them.

Why it matters: Splats have been photorealistic but frozen: the lighting from the moment of capture is welded into the data. Making them behave under DCC lights is the ecosystem's current frontier — SideFX shipped Houdini relighting nodes just days ago, and Herrera credits Houdini's self-shadowing as the prompt for building this.

Go deeper (1 min. read) ⟶
Tessa Marchetti

Tessa Marchetti

3 days ago-Industry

Esri puts georeferenced splats into Unreal Engine

A gaussian splat cloud anchored by a map pin to a gridded map plane

Esri has released ArcGIS Maps SDK for Unreal Engine 2.4.0 with a new gaussian splat layer type, letting developers stream georeferenced splat datasets from ArcGIS Online or ArcGIS Enterprise directly into Unreal scenes.

Why it matters: This completes Esri's splat rollout across its stack: ArcGIS Pro 3.6 introduced the layer, the JavaScript SDK 5.0 took it to the browser, ArcGIS Enterprise 12.1 brought it on-premises — and now it lands in the engine that runs simulations, digital twins and serious games.

Go deeper (1 min. read) ⟶
View all stories ⟶

The latest

  • Two smeared input photographs of a black car, beside the method's reconstructed view and the ground truth, with the grille crop enlarged underneath each

    Two blurry photos are now enough for a 3D scene

  • A query photo of a red handbag among clutter, beside a baseline extraction where the bag comes out surrounded by smeared background, beside Seed2GS's clean isolated bag on black

    Pull an object out of a splat you didn't capture

  • Two aerial views of the same reconstructed conifer forest: at ten simulated minutes a small orange and black burn scar, and at fifty minutes a black scar covering most of the frame

    They set fire to a gaussian splat of a real forest

  • Three columns: the raw compressive measurement as an unreadable scatter of speckle, the method's reconstruction of a plate of hot dogs and a vending machine, and the ground truth beside it

    One exposure in, a whole 3D scene out

  • The same orchard branch twice: on the left the radiance field rendered photorealistically, on the right the semantic field, with apples picked out in red and foliage in green

    Radiance fields stop being pictures and start being places

  • Three living-room scenes in a grid: the original capture, the ground-truth mask with one person filled in red, and the system's predicted mask filling the same person in blue

    Volumetric video turns privacy into a 3D problem

  • Two reconstructions of the same scene. In the TokenGS result the orange gaussians decoded from one token are scattered across distant parts of the scene; in the LocusGS result the same orange points sit in one tight cluster

    LocusGS gives instant splats a sense of place

  • A walnut with metal pins inserted, beside two CT reconstructions of it: the conventional FDK result streaked with bright artifacts, and the paper's result showing the same slices clean

    Splatting leaves graphics and turns up in the CT scanner

splats

Short daily briefs on the evolution of media — gaussian splats, volumetric video, dome theaters, headsets, and the research underneath.

Newsroom

  • Latest
  • All stories
  • RSS feed
© 2026 Splats · Terms · Privacy