Bake It Till You Make It

Ultrafast Spatial Texture-Atlas Splatting

Anonymous authors · ICLR 2027 submission · material added during the discussion phase (November 2026)

The supplementary demos of the submission are unchanged at the original page. This page holds only what was added in response to the reviews. No visitor tracking.

Where each question is addressed

All results on this page will be included in the revised manuscript.

Mipmapping of the texture atlas

Each bundle is the paper's model plus six box-filtered coarser levels of every surfel tile, down to one texel, with no retraining. The viewer opens with anisotropic mipmapping on: each pixel samples the atlas at the level of its own texel footprint, with up to 16 taps along the long axis. The Mipmaps toggle in the panel switches back to level 0, the renderer used for the paper's results. Orbit slowly over the floor or the lawn and toggle to compare. Needs WebGPU: Chrome or Edge 113+ on a laptop or desktop.

Added results

HD and 4K on an RTX 5090, frames per second

Held-out test cameras, SM clock locked at 2.4 GHz. CUDA = each method's own released renderer; HW = the same models through one hardware rasterizer (Vulkan), FastGS and Speedy-Splat with their own bounding boxes. Primitives: ours 63k / 85k / 66k / 149k, FastGS 207k / 253k / 252k / 661k, Speedy-Splat 119k / 257k / 314k / 535k (room / truck / drjohnson / garden). Garden ran at the unlocked clock and reads about 10 % high against the other three scenes.

sceneresolutionours CUDAFastGS CUDASpeedy CUDAours HWFastGS HWSpeedy HW
room1920×1279128711031169235314591167
room3840×2558469635622884485358
truck1920×10709901082103914721165899
truck3840×2141314630541462405296
drjohnson1920×1262112010651055265417061270
drjohnson3840×2525360544540895552401
garden1920×1244105293282619181108950
garden3840×2488384590504705507384

On the hardware rasterizer ours is fastest in every cell and the lead grows with resolution. On the software path our fragment is more expensive (ray/plane hit plus one texture fetch against one Gaussian evaluation), so at 4K the 3DGS-family renderers are ahead there.

Apple M3 Pro laptop (WebGPU), frames per second

sceneresolutionoursFastGSSpeedy-Splat
gardenevaluation297199228
roomevaluation317206256
truckevaluation338297338
drjohnsonevaluation442289328
garden1920 wide176137159
room1920 wide226145178
truck1920 wide123121146
drjohnson1920 wide241162192
garden3840 wide555259
room3840 wide664149
truck3840 wide353443
drjohnson3840 wide674555

MacBook Pro with an Apple M3 Pro, Chrome, WebGPU; FastGS and Speedy-Splat through the same WebGPU sort and blend, measured in the same session as ours. Primitives: ours 149k / 63k / 85k / 66k, FastGS 661k / 207k / 253k / 252k, Speedy-Splat 535k / 119k / 257k / 314k (garden / room / truck / drjohnson). On truck at 1920 and 3840 wide and on garden at 3840 wide our per-fragment cost (ray/plane hit and one texture fetch) outweighs our lower overdraw.