Enum Vae3dStep 

pub enum Vae3dStep<'a> {
    Conv {
        w: &'a [f32],
        b: &'a [f32],
        in_ch: usize,
        out_ch: usize,
    },
    NormSilu {
        eps: f32,
    },
    Save,
    AddSaved,
    DepthToSpace {
        p: (usize, usize, usize),
    },
}

One stage of the LTX-2.5 video VAE’s decode over a volume [c][f][h][w].

Variants

Conv

A 3x3x3 convolution with zero padding in space and replicated frames in time: w [out_ch][in_ch * 27], b [out_ch].

Fields

w: &'a [f32]

b: &'a [f32]

in_ch: usize

out_ch: usize

NormSilu

The RMS over the channels at every position (eps under the root), then SiLU.

Fields

eps: f32

Save

Keep the activation for a later AddSaved.

AddSaved

Add the kept activation.

DepthToSpace

Depth to space over (frames, height, width), the channels laid out (c p1 p2 p3); a frame stride of 2 leaves the first frame out.

Fields

p: (usize, usize, usize)