Struct VideoVaeEncoder 

pub struct VideoVaeEncoder { /* private fields */ }

The encoder half: [3, h, w] in [-1, 1] → normalized latents [24, h/16, w/16].

Implementations

impl VideoVaeEncoder

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

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

A single frame in [-1, 1], [3, h, w] → normalized latents [z, h/16, w/16], spatially tiled exactly as the reference does: the posterior MEAN, no sampling (what a video keyframe conditions on).

pub fn encode_frame_sampled( &self, rgb: &[f32], h: usize, w: usize, seed: u64, ) -> (Vec<f32>, usize, usize)

The same frame as a SAMPLE of the posterior — mean + σ·ε with logvar clamped to [-30, 20] — which is what MiniMax-H3-Image’s reference pictures condition on. The reference draws ε from a torch generator at a fixed seed; ours is its own stream of the same law, fixed by seed.