Struct DiffusionConfig 

pub struct DiffusionConfig {

    pub num_views: usize,
    pub guidance_scale: f64,
    pub num_inference_steps: usize,
    pub upsampler_steps: usize,
    pub image_size: usize,
    pub latent_size: usize,
    pub latent_channels: usize,
    pub unet_in_channels: usize,
    pub unet_out_channels: usize,
    pub cross_attention_dim: usize,
    pub clip_embed_dim: usize,
    pub time_embed_dim: usize,
    pub base_channels: usize,
    pub channel_mult: Vec<usize>,
    pub layers_per_block: usize,
    pub attention_head_dim: Vec<usize>,
    pub transformer_layers_per_block: Vec<usize>,
    pub norm_num_groups: usize,
    pub norm_eps: f64,
    pub camera_pose_dim: usize,
    pub use_linear_projection: bool,
    pub vae_scale_factor: f64,
    pub use_flash_attention: bool,
    pub flash_attention_block_size: usize,
    pub upsampler_mode: Option<UpsamplerMode>,
}

Full configuration for the multi-view diffusion model.

Contains all hyperparameters for the diffusion pipeline, including U-Net architecture, attention settings, CFG parameters, and optional upsampling.

Examples

use oxigaf_diffusion::DiffusionConfig;

// Use default configuration (256×256, guidance_scale=3.0)
let config = DiffusionConfig::default();

// Customize guidance scale for stronger conditioning
let mut config = DiffusionConfig::default();
config.guidance_scale = 7.5;

// Enable upsampling for 512×512 output
use oxigaf_diffusion::UpsamplerMode;
config.upsampler_mode = Some(UpsamplerMode::SdX2);

Fields

num_views: usize

Number of views to generate simultaneously (default: 4).

guidance_scale: f64

Classifier-free guidance scale for IP-Adapter conditioning (default: 3.0).

Controls the strength of reference image conditioning. Must be >= 1.0. Higher values increase identity preservation but may reduce diversity.

• 1.0: No guidance (pure conditional)
• 3.0-7.5: Balanced (recommended)
• >10.0: Strong conditioning (may oversaturate)

num_inference_steps: usize

Number of DDIM denoising steps (default: 50).

upsampler_steps: usize

Number of latent upsampler denoising steps (default: 10).

image_size: usize

Input/output image resolution before upscaling (default: 256).

latent_size: usize

Latent spatial size (image_size / 8).

latent_channels: usize

Number of latent channels produced by the VAE (default: 4).

unet_in_channels: usize

U-Net input channels: latent_channels + normal-map latent channels (default: 8).

unet_out_channels: usize

U-Net output channels (default: 4).

cross_attention_dim: usize

Cross-attention dimension (SD 2.1 = 1024).

clip_embed_dim: usize

CLIP image embedding dimension (ViT-H/14 = 1280).

time_embed_dim: usize

Time embedding dimension (default: 1280).

base_channels: usize

Base channels for the U-Net (default: 320).

channel_mult: Vec<usize>

Channel multipliers per U-Net stage.

layers_per_block: usize

Layers per block in the U-Net.

attention_head_dim: Vec<usize>

Number of attention heads per head-dim for each stage.

transformer_layers_per_block: Vec<usize>

Number of transformer blocks per attention stage.

norm_num_groups: usize

Group-norm number of groups (default: 32).

norm_eps: f64

Group-norm epsilon.

camera_pose_dim: usize

Camera pose input dimension (4×3 flattened = 12).

use_linear_projection: bool

Whether to use linear projection in spatial transformer.

vae_scale_factor: f64

VAE scaling factor for latent space.

use_flash_attention: bool

Whether to use flash attention for memory-efficient O(N) attention. When enabled, uses block-wise computation with online softmax. Falls back to standard O(N^2) attention when disabled. Default: true (when feature is enabled).

flash_attention_block_size: usize

Block size for flash attention tiled computation. Larger blocks use more memory but may be faster due to better cache utilization. Default: 64.

upsampler_mode: Option<UpsamplerMode>

Upsampler mode for latent upsampling (32×32 → 64×64).

• None: No upsampling, output is 256×256
• Some(SdX2): Use sd-x2-latent-upscaler, output is 512×512
• Some(BilinearVae): Use bilinear upsampling, output is 512×512

Default: None (256×256 output).

Implementations

impl DiffusionConfig

pub fn stage_channels(&self, stage: usize) -> usize

Channel count for a given U-Net stage index.

pub fn num_stages(&self) -> usize

Total number of U-Net stages.

Trait Implementations

impl Clone for DiffusionConfig

fn clone(&self) -> DiffusionConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for DiffusionConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for DiffusionConfig

fn default() -> Self

Returns the “default value” for a type.