burn_reconstruction 🕊️🔥🌌

bevy / burn feed-forward gaussian splatting. view the wasm example

features

• multi-view images -> 3d gaussians (YoNoSplat path)
• multi-model API/CLI/wasm/bevy selection surface (YoNoSplat + ZipSplat)
• CLI export to GLB (KHR_gaussian_splatting)
• bevy UI (native + wasm)
• model auto-bootstrap + cache (parts-first burnpack)
• native Burn/Rust ZipSplat forward path (DA3/DINO backbone + ZipSplat fusion/head)
• official ZipSplat checkpoint auto-cache/import to native burnpack
• ZipSplat upstream-output numerical parity fixture/test path
• bevy_synth integration
• bevy_zeroverse fine-tuning/training

setup

install

# cli
cargo install burn_reconstruction

# bevy app
cargo install bevy_reconstruction

usage

burn_reconstruction \
  --images view0.png view1.png view2.png \
  --output /tmp/scene.glb

burn_reconstruction \
  --model zipsplat \
  --weights-format bpk \
  --quality compact \
  --zipsplat-r 4 \
  --images view0.png view1.png view2.png \
  --output /tmp/scene_zipsplat.glb

ZipSplat runs through the native Burn/Rust pipeline. When --zipsplat-weights is omitted on native targets, the CLI downloads the official ZipSplat checkpoint to ~/.burn_reconstruction/models/zipsplat, imports it to zipsplat.bpk, converts the requested precision (zipsplat_f16.bpk by default), and reuses that cache on later runs. To import manually:

cargo run -p burn_reconstruction --bin import -- \
  --model zipsplat \
  --zipsplat-weights ~/.burn_reconstruction/models/zipsplat/zipsplat-da3g-252p.tar \
  --zipsplat-output ~/.burn_reconstruction/models/zipsplat/zipsplat \
  --precision both \
  --parts true

Useful native ZipSplat flags:

• --weights-format bpk|safetensors selects the native checkpoint format.
• --zipsplat-weights <path> selects an explicit converted ZipSplat checkpoint instead of auto-cache.
• --zipsplat-r <N> controls the retained-token reduction factor; higher values emit fewer Gaussians.

ZipSplat numerical parity uses the official upstream PyTorch implementation only as offline reference-export tooling:

git clone --depth 1 https://github.com/cvg/ZipSplat /tmp/ZipSplat
python3 -m venv .venv-zipsplat-ref
.venv-zipsplat-ref/bin/python -m pip install torch safetensors einops numpy pillow
.venv-zipsplat-ref/bin/python tool/export_zipsplat_reference.py
cargo test -p burn_reconstruction --features correctness --test parity -- --nocapture zipsplat

The runtime ZipSplat path remains native Burn/Rust and does not call Python.

bevy_reconstruction -- \
  --image assets/images/re10k/0.png \
  --image assets/images/re10k/1.png \
  --image assets/images/re10k/2.png

note, input images are optional for bevy_reconstruction ZipSplat browser/wasm inference requires hosted zipsplat_f16.bpk.parts.json and matching parts.

license

licensed under either of:

• Apache License, Version 2.0
• MIT license

at your option.

note: model weights have their own license

references

• bevy_gaussian_splatting
• burn_dino
• YoNoSplat
• ZipSplat