Struct MultiViewUNet 

pub struct MultiViewUNet { /* private fields */ }

The multi-view U-Net for diffusion-based avatar generation.

Architecture matches SD 2.1 but with multi-view cross-attention in every spatial transformer block, camera-pose conditioning added to the timestep embedding, and IP-adapter cross-attention for reference-image conditioning.

Implementations

impl MultiViewUNet

pub fn new(vs: VarBuilder<'_>, config: &DiffusionConfig) -> Result<Self>

Build the U-Net from a DiffusionConfig and VarBuilder.

pub fn forward( &self, sample: &Tensor, timestep: usize, context: Option<&Tensor>, camera_poses: Option<&Tensor>, ip_tokens: Option<&Tensor>, ) -> Result<Tensor, DiffusionError>

Forward pass.

• sample: (B*V, in_channels, H, W) noisy latent input.
• timestep: scalar timestep (will be broadcast to batch).
• context: (B*V, seq_len, cross_attn_dim) text/null embedding.
• camera_poses: (B*V, pose_dim) flattened extrinsics.
• ip_tokens: (B*V, ip_len, ip_dim) CLIP image tokens.

Errors

Returns DiffusionError::SkipConnectionUnderflow if skip connections are exhausted before all up blocks have consumed them.

Trait Implementations

impl Debug for MultiViewUNet

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

Formats the value using the given formatter.