Struct LatentUpsampler 

pub struct LatentUpsampler { /* private fields */ }

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

This enables 512×512 output resolution (vs 256×256) by upsampling the latent representation before VAE decoding. The upsampler uses a separate U-Net model with 10-step DDIM denoising in latent space.

Modes

• SdX2: Uses the sd-x2-latent-upscaler U-Net with DDIM denoising.
• BilinearVae: Fallback mode that uses simple bilinear upsampling.

Example

use oxigaf_diffusion::{LatentUpsampler, UpsamplerMode};
use candle_core::Device;

let device = Device::Cpu;
let upsampler = LatentUpsampler::load(
    UpsamplerMode::SdX2,
    "weights/upsampler",
    &device,
)?;

// Upsample latents from 32×32 to 64×64
let upsampled = upsampler.upsample(&latents, 10)?;

Implementations

impl LatentUpsampler

pub fn load( mode: UpsamplerMode, weights_path: &Path, device: &Device, ) -> Result<Self, DiffusionError>

Load a latent upsampler from weights.

Arguments

• mode - Upsampler mode (SdX2 or BilinearVae)
• weights_path - Path to upsampler weights directory (for SdX2 mode)
• device - Device to load the model on

Errors

Returns DiffusionError::ModelLoad if weight loading fails.

pub fn upsample( &mut self, latents: &Tensor, num_steps: usize, ) -> Result<Tensor, DiffusionError>

Upsample latents from 32×32 to 64×64.

Arguments

• latents - Input latents (B, 4, 32, 32)
• num_steps - Number of DDIM denoising steps (typically 10)

Returns

Upsampled latents (B, 4, 64, 64)

Errors

Returns DiffusionError::Inference if upsampling fails.

Trait Implementations

impl Debug for LatentUpsampler

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

Formats the value using the given formatter.