Struct ImageSession 

pub struct ImageSession { /* private fields */ }

A loaded, resident text-to-image pipeline. Build once with Self::load, then call Self::render per prompt.

Implementations

impl ImageSession

pub fn load( dit_gguf: &Path, vae_weights: &Path, te_source: &TeSource, tokenizer_dir: &Path, ) -> Result<Self, PipelineError>

Load every model asset and keep it resident.

te_source selects the 4-bit MLX safetensors or the f32 .npy dir, same as crate::pipeline::TextToImageCfg. The text encoder is put in resident mode so its dequantised weights persist across renders.

Errors

PipelineError wrapping whichever asset failed to load.

pub fn warm(&self) -> Result<Duration, PipelineError>

Populate the resident text-encoder cache so the first real render runs at warm speed. A single short encode touches every layer’s weights, so one pass dequantises the whole encoder. Returns how long warming took.

DiT and VAE weights are memory-mapped and page in lazily; this only warms the encoder, which is the dominant amortizable (re-dequant) cost.

Errors

PipelineError if the warm-up encode fails.

pub fn render( &self, params: &RenderParams, ) -> Result<RenderOutcome, PipelineError>

Render one image. Reuses the resident weights; nothing is re-loaded.

Errors

PipelineError wrapping whichever stage failed.