A major leap in digital humans for the COOLJAPAN ecosystem.
On February 24 we released OxiGAF 0.1.0 — a complete pure-Rust implementation of Gaussian Avatar Reconstruction from monocular videos using multi-view diffusion.
This is the first production-ready Rust implementation of the GAF method (arXiv:2412.10209) that turns a casual selfie video into a high-fidelity, animatable 3D Gaussian avatar — all without multi-camera rigs, heavy Python dependencies, or unsafe code.
Why OxiGAF is revolutionary
Traditional avatar reconstruction (e.g., Gaussian Splatting papers, NeRF-based methods) relies on:
- Multi-view capture setups
- Python + PyTorch + CUDA-heavy stacks
- Slow training (hours to days)
OxiGAF changes the game:
- Single monocular video input only
- 100% Pure Rust (no C/C++/Fortran)
- Differentiable rendering via wgpu (GPU on every platform)
- Real-time inference capable
- Seamless integration with oxihuman (parametric body) and oximedia (video pipelines)
Technical Deep Dive: The GAF Pipeline in Rust
The core innovation combines three pillars:
-
Multi-View Diffusion (
oxigaf-diffusion)- UNet-based diffusion model that generates 4–8 consistent views from a single input frame.
- IP-Adapter for identity preservation + classifier-free guidance (scale 1.0–20.0).
- Cross-view attention + explicit camera pose embedding.
- Latent upsampler (32→64) for high-quality 512×512 outputs.
-
Differentiable Gaussian Splatting (
oxigaf-render)- CPU reference rasterizer + full GPU implementation via wgpu.
- Gradient-verified 35 tests (relative error < 1e-3).
- Supports opacity, spherical harmonics (SH degree 3), covariance scaling, and rotation.
-
FLAME Parametric Binding (
oxigaf-flame)- Linear Blend Skinning (LBS) + normal maps.
- Safetensors I/O + video sequence caching with LRU.
- Bidirectional PyTorch ↔ OxiGAF weight conversion via
oxigaf-bridge.
The full training loop verifies gradients end-to-end, ensuring the entire pipeline is differentiable and optimizable.
What’s inside 0.1.0
- Full GAF pipeline (diffusion → Gaussian optimization → FLAME binding)
- 796 passing tests (100% coverage)
- Feature flags:
cuda,metal,simd,flash_attention,mixed_precision - CLI + Python bindings ready
- Zero
unwrap()in critical paths
This is the foundation
OxiGAF is now the avatar reconstruction layer for the entire COOLJAPAN stack:
- oxihuman + OxiGAF = parametric body + photorealistic face & hair from video
- oximedia pipes reconstructed avatars into real-time video streams
- mielinOS agents get photorealistic digital twins
- oxilean can formally verify avatar physics and animation constraints
Repository: https://github.com/cool-japan/oxigaf
Star the repo if you want digital humans that are fast, safe, and truly sovereign.
The era of “upload to a cloud service for avatar reconstruction” is over.
Pure Rust Gaussian avatars are here — and they run everywhere.
— KitaSan at COOLJAPAN OÜ
February 24, 2026