Crate oxigaf_diffusion 

oxigaf-diffusion

Multi-view diffusion model inference for GAF.

Implements the full pipeline: CLIP image encoding → multi-view U-Net denoising with camera-conditioned cross-view attention → VAE decoding.

Cargo Features

This crate supports the following feature flags:

• default = ["accelerate", "flash_attention"]: Default features for CPU-only inference with optimizations

• accelerate: Uses platform-native BLAS/LAPACK for CPU tensor operations

  • macOS: Accelerate framework
  • Linux: OpenBLAS or Intel MKL

• cuda (platform-specific): Enables NVIDIA GPU acceleration via CUDA

  • Requires CUDA toolkit installed
  • Not available on macOS

• metal (platform-specific): Enables Apple Silicon GPU acceleration via Metal

  • macOS only
  • Optimized for M1/M2/M3 chips

• flash_attention (enabled by default): Memory-efficient attention with O(N) complexity instead of O(N²)

  • Reduces memory usage by 2-4× for large images
  • Tiled computation for better cache locality

• mixed_precision (planned, not yet implemented): FP16/BF16 inference for reduced memory usage

  • Faster on GPUs with Tensor Cores
  • Lower memory footprint

Example usage:

# In Cargo.toml
# For CPU-only with flash attention
oxigaf-diffusion = { version = "0.1", default-features = true }

# For Apple Silicon with Metal acceleration
oxigaf-diffusion = { version = "0.1", features = ["metal", "flash_attention"] }

# For NVIDIA GPU with CUDA
oxigaf-diffusion = { version = "0.1", features = ["cuda", "flash_attention"] }

Re-exports

pub use clip::ClipImageEncoder;

pub use config::DiffusionConfig;

pub use pipeline::MultiViewDiffusionPipeline;

pub use pipeline::MultiViewOutput;

pub use scheduler::DdimScheduler;

pub use scheduler::PredictionType;

pub use unet::MultiViewUNet;

pub use upsampler::LatentUpsampler;

pub use upsampler::UpsamplerMode;

pub use vae::Vae;

pub use flash_attention::flash_attention;

pub use flash_attention::flash_attention_with_config;

pub use flash_attention::FlashAttention;

pub use flash_attention::FlashAttentionConfig;

Modules

attention

Attention-based building blocks for multi-view diffusion.

camera

Camera-pose conditioning MLP.

clip

CLIP image encoder for reference-image conditioning.

config

Configuration for the multi-view diffusion pipeline.

flash_attention

Flash Attention: Memory-efficient attention with block-wise computation.

pipeline

Full multi-view diffusion pipeline.

scheduler

DDIM scheduler with v-prediction parameterisation.

unet

Multi-view U-Net with camera-conditioned cross-view attention.

upsampler

Latent upsampler for 32×32 → 64×64 latent upsampling.

vae

Variational Autoencoder (SD 2.1 compatible).

Enums

DiffusionError

Errors that can occur during diffusion model operations.

Type Aliases

DiffusionResult

Result type for diffusion operations.