loading models

colossus

Virtualized geometry from scratch, as in Nanite: every cluster of every statue picks the coarsest detail within a pixel of the original. Below, a lighter live version, then frames from the native Vulkan viewer on an RX 9070 XT.

900 instances of Lucy and the XYZ RGB dragon in sunlight, shadowed by virtual shadow maps
900 statues of Lucy and the XYZ RGB dragon: 15.9 billion triangles at full detail, drawn at 1920x1080 in 1.48 ms with soft shadows, bounce light, ambient occlusion and antialiasing. About 4.8 million reach the screen.
the crowd coloured by LOD level, blue for full detail through red to magenta for the coarsest
LOD levels, blue for full detail through red to magenta for the coarsest. Near dragons draw mid levels, distant statues the coarsest, and single statues mix levels as they recede.
Lucy, each cluster in its own colour
Lucy, each cluster of up to 128 triangles in its own colour. Clusters are grouped, simplified and grouped again into a hierarchy where any mix of levels joins without cracks.
the crowd coloured by rasterizer: orange for compute, blue for hardware
Which rasterizer drew each pixel: orange a compute rasterizer, blue mesh shaders. Pixel-sized triangles are what hardware handles worst, so only the nearest surfaces get it.
Lucy in each material: veined marble, layered sandstone, patinated bronze, hammered gold, speckled granite
Lucy and the dragon have no UVs, so their materials come from noise in each model's own space: veined marble, layered sandstone, bronze greening in its folds, hammered gold, granite.
Horatio Greenough's George Washington, scanned by the Smithsonian, close up in its own texture
Horatio Greenough's George Washington (1840), the Smithsonian American Art Museum's scan: 17 million triangles and its own 4096x4096 texture, streamed in tiles beside the geometry.
a crowd of running, walking and watchful foxes, each skinned and animated
900 Khronos foxes, skinned and animated, each subdivided to 590 thousand triangles and still picking its level per cluster: errors are measured in every pose of their animations.