Architecture

worldgen/worldgen_t2s.py     prompt -> panorama -> depth -> scene
worldgen/worldgen_i2s.py     image -> Fill panorama -> depth -> scene
worldgen/worldgen_i2s_prep.py 90-degree band cond/mask
worldgen/worldgen_da2.py     ncnn Vulkan DA-2 bridge

tools/stable-diffusion.cpp   FLUX.1-dev / FLUX.1-Fill-dev (C++/Vulkan)
tools/da2-ncnn/da2ncnn.exe   DA-2 (C++/ncnn/Vulkan)
rust/                        worldgen-vk (Rust): RGBD -> 3DGS/mesh PLY
da2/main.cpp                 ncnn runner source
patches/da2/                 DA-2 export fixes

FLUX stage

stable-diffusion.cpp with SD_VULKAN=ON. Diffusion, T5, CLIP and VAE all run on Vulkan0 (the VAE uses --vae-tiling --vae-tile-size 512x512 to avoid a 9 s untiled decode).

DA-2 stage

DA-2 is a DINOv2 ViT-L/14 + SphereViT head. The export path:

Apply patches/da2/ (remove quantile, separate QKV projections).
Trace at 1x3x512x1024.
pnnx -> ncnn (0 custom layers).
da2ncnn.exe runs the graph on Vulkan.

The runner must be built with -DNCNN_VULKAN and call net.set_vulkan_device(0); otherwise ncnn silently uses CPU (8.7 s vs 3.05 s).

Rust stage

worldgen-vk both converts pano+depth to:

standard 17-float 3DGS PLY
vertex-colored mesh PLY with a panorama UV seam fix

It is CPU geometry, 0.29 s for 524,288 splats / 1,046,528 triangles.

Coordinate conventions

WorldGen uses the same normalized camera/center formula as the Python reference; validation of center/scale/quaternion is ~1e-7.