Struct MultiViewDiffusionPipeline 

pub struct MultiViewDiffusionPipeline { /* private fields */ }

The full multi-view diffusion pipeline.

Implementations

impl MultiViewDiffusionPipeline

pub fn load( config: DiffusionConfig, weights_dir: &Path, device: &Device, ) -> Result<Self, DiffusionError>

Load a pipeline from a directory of safetensors files.

Expected files:

• unet/diffusion_pytorch_model.safetensors
• vae/diffusion_pytorch_model.safetensors
• image_encoder/model.safetensors
• upsampler/diffusion_pytorch_model.safetensors (optional, for SdX2 mode)

pub fn generate( &mut self, reference_image: &Tensor, normal_map_latents: &Tensor, camera_poses: &Tensor, _seed: u64, ) -> Result<MultiViewOutput, DiffusionError>

Generate multi-view images from a reference image and camera poses.

• reference_image: (1, 3, 224, 224) normalised image for CLIP.
• normal_map_latents: (num_views, latent_channels, h, w) encoded normal maps.
• camera_poses: (num_views, pose_dim) flattened extrinsics per view.
• seed: RNG seed for reproducibility.

Classifier-Free Guidance (CFG)

This pipeline implements CFG for IP-Adapter conditioning:

• Conditional pass: Uses IP tokens from CLIP-encoded reference image
• Unconditional pass: Skips IP tokens (no reference conditioning)
• Formula: pred = uncond + guidance_scale * (cond - uncond)

The guidance_scale parameter (from config) controls the strength of conditioning. Typical values:

• 1.0 = no guidance (unconditional generation)
• 3.0-7.5 = balanced (default: 3.0 for GAF)
• >10.0 = strong conditioning (may oversaturate)

Errors

Returns DiffusionError::Inference if guidance_scale < 1.0 or if any tensor operation fails during generation.