Struct VaeDecoder 

pub struct VaeDecoder {
    pub latent_channels: usize,
    pub scaling_factor: f32,
    pub shift_factor: f32,
    /* private fields */
}

The full decoder: conv_in → mid(res, attn, res) → up-blocks → GroupNorm + SiLU → conv_out, plus the diffusers latent de-normalization z/scaling_factor + shift_factor.

Fields

latent_channels: usize

scaling_factor: f32

shift_factor: f32

Implementations

impl VaeDecoder

pub fn chain_args(&self) -> VaeChainArgs<'_>

The borrowed device-chain view (see VaeChainArgs).

pub fn decode_fast(&self, z: &[f32], h: usize, w: usize) -> Vec<f32>

decode with the resident device chain first: de-normalise on the host, try gpu::vae_decode_chain, and fall back to decode (the unchanged Lumina path) when the backend declines. Same contract as decode: model-scale latents in, RGB [3, 8h, 8w] in ≈[-1, 1] out.

impl VaeDecoder

pub fn load_dir(dir: &Path) -> Result<Self, String>

Load from a diffusers vae/ directory (config.json + diffusion_pytorch_model.safetensors).

pub fn from_cmf(model: &CmfModel) -> Result<Self, String>

Load from a packaged imagegen .cmf (vae.* tensors, stored f16/f32, + vae.config_json).

pub fn decode(&self, z: &[f32], h: usize, w: usize) -> Vec<f32>

Decode latents [latent_channels, h, w] (model scale, i.e. as produced by the diffusion loop) into RGB [3, 8h, 8w] in [-1, 1].