Status: Exploration

@luma.gl/three is a private, version-pinned prototype. Select WebGPU or WebGL2 to compare the luma backend with upstream Three.js on separate canvases/devices. Orbit, pan and zoom in either pane synchronize both views.

The selector covers textured glTF PBR, directional PCF shadows, morph animation, skinning, instancing, physical materials, transmission and animated TSL graphics. WebGPU also validates compute readback. WebGL2 compute is explicitly unsupported; transform-feedback translation is deferred. The mixed Three/luma object adapter scene currently requires WebGPU.

This experiment does not promise general Three.js compatibility. See the adapter documentation for support and ownership boundaries.

Shared buffer / luma compute explores a complementary thin bridge: luma compute generates 4,225 mesh vertices on a device shared with upstream WebGPURenderer. Three renders that same GPU buffer directly, with no vertex readback or re-upload. The other pane uses upstream TSL displacement as a synchronized reference. This scene requires WebGPU and uses private, version-pinned attribute hooks.

Three meshes inside deck.gl puts a Three cube and a native deck ScatterplotLayer disk in the same WebGPU render pass, sharing the device and depth attachment. Both panes use a synchronized camera; the right pane is an upstream Three reference. Reverse the deck layer order to check depth occlusion, or pause the animation to inspect it. This first integration supports Cartesian views and untextured basic materials; geospatial projection and picking are deferred.

Lantern courtyard inside deck.gl / PBR retains the glTF model’s original standard material, textures, directional PCF shadows and environment lighting. Native deck polygon tiles, survey rings, animated columns and markers surround it in the same shared-depth pass. The right pane renders matching objects with upstream Three.js. This scene requires WebGPU and uses a prepared scene adapter.

Geographic Lanterns + picking places three textured glTF instances in a local meter frame around a San Francisco coordinate, with altitude, scale and heading. Native deck buildings and ground share depth. Drag or zoom the left map to update both views; click either pane for a stable selection ID and matching highlight. Toggle the model layer, reverse draw order, resize, or dispose/recreate the scene. Picking uses CPU triangle raycasting with explicit building occluders; general deck GPU picking of Three meshes and large geographic extents remain deferred.

Buildings act as occluders for CPU selection; only Lantern instances receive selection IDs and highlights. Their bases are normalized from geometry bounds before applying instance scale, so placement altitude describes the model base.