Struct VaeDecoderConfig 

pub struct VaeDecoderConfig {
    pub out_channels: usize,
    pub latent_channels: usize,
    pub block_out_channels: Vec<usize>,
    pub layers_per_block: usize,
    pub norm_num_groups: usize,
    pub sample_size: usize,
    pub scaling_factor: f64,
}

Frozen config for the Stable-Diffusion VAE decoder.

Mirrors the decoder-relevant subset of AutoencoderKL.config for runwayml/stable-diffusion-v1-5. The defaults match SD 1.5 exactly.

Fields

out_channels: usize

Number of input channels of the image the encoder consumes (and therefore of the image the decoder produces). For SD 1.5: 3.

latent_channels: usize

Number of latent channels. For SD 1.5: 4.

block_out_channels: Vec<usize>

Per-block-level output channel counts (in encoder order: from the highest-resolution block out). For SD 1.5: [128, 256, 512, 512]. The decoder walks these in reverse, so the first block after conv_in has block_out_channels[-1] channels (= 512).

layers_per_block: usize

Number of resnet layers in each Encoder / Decoder up- or down-block. The decoder’s UpDecoderBlock2D uses layers_per_block + 1 resnets (the diffusers convention). For SD 1.5: 2 (so each up-block has 3 resnets).

norm_num_groups: usize

Number of GroupNorm groups (decoder-internal norm1 / norm2 / conv_norm_out). For SD 1.5: 32.

sample_size: usize

Spatial size the encoder accepts (and the decoder produces). For SD 1.5: 512.

scaling_factor: f64

VAE latent scaling factor. The decoder pre-divides the latent by this value (matching AutoencoderKL.decode). For SD 1.5: 0.18215.

Implementations

impl VaeDecoderConfig

pub fn sd_v1_5() -> Self

SD 1.5 VAE decoder config (alias for Default::default()).

pub fn validate(&self) -> FerrotorchResult<()>

Validate field bounds (positive sizes, channels divisible by norm_num_groups, at least one resolution).

Errors

Returns FerrotorchError::InvalidArgument for any out-of-bounds or arithmetic-incompatible field.

pub fn resnets_per_up_block(&self) -> usize

Number of resnets in each UpDecoderBlock2D (the diffusers convention is layers_per_block + 1).

pub fn num_up_blocks(&self) -> usize

Number of up-blocks (== number of down-blocks the encoder used, == block_out_channels.len()).

pub fn from_json_str(s: &str) -> FerrotorchResult<Self>

Parse a vae/config.json document into a VaeDecoderConfig.

Recognised keys (all optional — anything missing falls back to the SD-1.5 defaults):

• out_channels, latent_channels, block_out_channels, layers_per_block, norm_num_groups, sample_size, scaling_factor.

Errors

Returns FerrotorchError::InvalidArgument on malformed JSON or a wrong-type field (e.g. block_out_channels not an array of integers).

pub fn from_file(path: &Path) -> FerrotorchResult<Self>

Parse a vae/config.json file from disk.

Errors

Returns FerrotorchError::InvalidArgument for I/O or parse failures (file missing, malformed JSON, wrong-type field).

Trait Implementations

impl Clone for VaeDecoderConfig

fn clone(&self) -> VaeDecoderConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for VaeDecoderConfig

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

Formats the value using the given formatter.

impl Default for VaeDecoderConfig

fn default() -> Self

Returns the “default value” for a type.