Struct SamRunner 

pub struct SamRunner {
    pub arch: SamArch,
    pub weights: PathBuf,
    pub device: Device,
    pub config_path: Option<PathBuf>,
}

SAM runner — owns the resolved config and dispatches the per-arch forward pass. SAM 1 / 2 / 3 differ enough in their prompting that we keep the heavy result type (SamPredictionAny) as a discriminated union the caller matches on.

Fields

arch: SamArch

weights: PathBuf

device: Device

config_path: Option<PathBuf>

Implementations

impl SamRunner

pub fn builder(arch: SamArch) -> SamRunnerBuilder

pub fn summary(&self) -> String

Print a human-readable summary — what the CLI prints before any per-arch image processing.

pub fn predict_image( &self, rgb: &[u8], h_in: usize, w_in: usize, points: Option<(&[f32], &[f32])>, boxes: Option<&[f32]>, text_tokens: &[u32], ) -> Result<SamPredictionAny>

End-to-end forward: image bytes → masks. Dispatches to the right per-arch entrypoint:

• SAM 1 → Sam::forward (multimask = true)
• SAM 2 → Sam2::predict_image (multimask = true)
• SAM 3 → Sam3::predict_image_text with the supplied text_tokens (required for SAM 3 — its decoder is text-conditioned). Pass an empty slice for arches that don’t use it.

rgb is HWC u8; points is (xy_pairs, labels) with one label per (x, y) pair (1 = foreground, 0 = background).

SAM-arch-specific defaults applied:

• cfg derived from environment variables (RLX_SAM_VARIANT for v1: vit_b/l/h; RLX_SAM2_VARIANT for v2: tiny/small/ base_plus/large); falls back to the smallest variant.
• multimask_output = true for v1 + v2.
• SAM 3 vit defaults to base.