Blender 5.3 Brings Gaussian Splats Home

For those who don’t know the tool: Blender is the Blender Foundation’s open-source DCC for modelling, animation, simulation, shading, rendering, compositing and increasingly substantial pipeline work. Blender 5.3 brings Gaussian splats into that workshop as native PointCloud data: captured scenes can arrive through PLY, SPZ or USD, pass into Geometry Nodes, and render directly through Workbench, Eevee Render and Cycles. In other words, the splats are no longer visiting Blender through the add-on entrance. They are getting their own key.

Blender 5.3 is putting 3D Gaussian Splats directly into Blender. That sentence is more interesting than it would have sounded two years ago, when splats were still largely something you admired in research demos, specialist viewers and slightly excitable browser tabs. With 5.3, Blender is turning them into a native scene-data type with import, Geometry Nodes access and rendering across its own engines.

The release is currently in alpha and expected on 10 November 2026. More interestingly, the splat work is not sitting off to one side as a novelty importer. Blender is wiring it into the application deeply enough that the little translucent blobs can start behaving like citizens rather than tourists.

The splats have moved in

Blender represents Gaussian splats as PointCloud data, with a new PointCloud type specifically for 3D Gaussian Splats. The data-block carries the scale, rotation, opacity and radiance information needed to reproduce the representation rather than flattening it into ordinary points. That immediately changes the character of the workflow. You are no longer asking an add-on to draw a strange external data format somewhere inside Blender. Blender itself knows what the thing is.

For artists coming from scanning and reconstruction workflows, this is a rather nice moment. A photographic environment reconstructed as a Gaussian splat can now enter the same DCC where cameras, characters, props, lights, animation and render settings already live. The splat does not suddenly become a beautifully retopologised hero asset with sensible UVs and a lovingly named material stack. That would be witchcraft. But it does become part of the Blender scene.

The official Blender 5.3 rendering notes already include useful images of the implementation, including Blender’s default splat shader and an example of a relit Gaussian Splat.

PLY walks straight in

PLY import can recognise Gaussian-splat data from its geometry and attributes. Blender then creates the appropriate point-cloud representation instead of treating the file as an ordinary mesh.

That sounds like small plumbing until you compare it with the previous reality of splat workflows, which often involved installing a dedicated add-on, learning its own object conventions and hoping the next Blender update did not turn the plug-in folder into an archaeological site.

SPZ also gets native import. Blender supports the currently defined format versions through SPZ 4 and converts the expected coordinate system into Blender’s own orientation during import.

So the front door now reads PLY and SPZ, and Blender knows why these files look like somebody spilled several million semi-transparent cornflakes across three dimensions.

There is also a USD route. Blender 5.3 imports the UsdVolParticleField3DGaussianSplat schema, giving facilities already using USD scene interchange another way of getting splat data into Blender. This is particularly welcome because Gaussian splats become much more interesting once they stop being isolated demo objects and start sharing scenes with the rest of production.

Nodes meet blobs

The more Blender-ish part begins with Geometry Nodes. Blender 5.3 adds an Import SPZ node and a Set Point Cloud Type node. The existing Import PLY node can identify Gaussian-splat data and output it as a point cloud. That gives procedural setups direct access to the new data type instead of forcing everything through a one-off manual import. And this is probably where many TDs will start smiling.

Once an asset type reaches Geometry Nodes, people inevitably begin asking unreasonable questions of it. Can we instance it? Filter it? Build ingest tools around it? Mix it with scene assembly? Turn Tuesday’s location scan into Thursday’s procedural contraption with seventeen exposed parameters and one control labelled “Do Not Touch”? Blender’s implementation does not magically answer every one of those questions (yet), but native Geometry Nodes access means the questions can now be asked inside the application itself.

EEVEE, Cycles and the gang

Workbench, EEVEE and Cycles can all render the new Gaussian-splat data. By default, Blender renders a splat using an emissive shader. The resolved radiance is available as an attribute, and the developer notes also show an early relighting route by replacing the default emission with a hard-surface BSDF such as Diffuse BSDF.

For a native first implementation, that is a spread. Workbench gives artists an immediate viewport route, EEVEE brings the splats into Blender’s realtime renderer, and Cycles places them inside the path-traced side of the house.

It also creates a more interesting mixed-scene proposition. A splat can sit alongside ordinary Blender geometry instead of demanding that the whole project relocate into a dedicated viewer. Cameras stay cameras. Lights stay lights. Your mesh is still a mesh. The photogrammetric-looking cloud of tiny ellipsoids is still a photogrammetric-looking cloud of tiny ellipsoids.

A little alpha seasoning

This is still Blender 5.3 alpha, so the current implementation has a few sharp corners. Blender documents that performance still needs work, colour handling is not yet fully correct between typical sRGB-trained splats and Blender’s linear rendering, and Apply Transform does not yet correctly update all of the new Gaussian attributes. Export is also not part of the present implementation.

There. Caveats delivered. Nobody needs to spend the rest of the article wearing a hard hat. The more useful perspective is that all of these issues now belong to Blender’s own implementation. The groundwork is in the core application, and the Gaussian-splat work was still reported as on track for 5.3 in Blender’s September project update.

November has plans

Gaussian splats are also arriving during a rather busy Blender development period. The 5.3 release notes already contain a new optional project system that lets Blender understand a project root beyond the currently open .blend file. The compositor is moving to an effects stack and gaining playback caching. Vulkan becomes the default backend on Windows, except ARM, and Linux. Workbench gains hardware-raytraced shadows on supported Metal and Vulkan hardware. Elsewhere, the 5.3 notes continue filling with smaller changes across animation, sculpting, assets, I/O and the interface.

Not all of these features are Gaussian-splat related, and they should not be shoved together into one giant marketing smoothie. They do, however, make the November release increasingly interesting. Blender’s own 2026 development overview had 5.3 pencilled in for November from the start, alongside continuing work on EEVEE, Cycles, Grease Pencil, editor workflows and production infrastructure. Watching the actual 5.3 branch fill in now gives that roadmap a little more weight.

There is a sense of several long-running threads meeting at once: more serious project awareness, more capable realtime graphics, more pipeline interchange, more compositor work, and now a scene representation that only recently required specialised tools becoming part of Blender proper.

Point releases have a habit of doing this. The headline may be one shiny feature, but the release becomes interesting because six other things have been growing in the garden while nobody was looking.

https://www.blender.org/releases/blender-5-3/

https://developer.blender.org/docs/release_notes/5.3/rendering/