Struct Sam2 

pub struct Sam2 { /* private fields */ }

Full SAM 2 model — owns the compiled image encoder + every host-side weight bundle. The encoder result is recomputed per call (no encoder-caching here; layer above can wrap if needed).

Implementations

impl Sam2

pub fn from_safetensors( weights_path: &str, cfg: Sam2Config, ) -> Result<Sam2, Error>

Load every SAM 2 component from a safetensors checkpoint and compile the image encoder for the CPU backend. For GPU/Metal, see Sam2::from_safetensors_on.

pub fn from_safetensors_on( weights_path: &str, cfg: Sam2Config, device: Device, ) -> Result<Sam2, Error>

Same as Sam2::from_safetensors but compiles the image encoder for the given backend. The cross-backend feature flags match SAM v1’s rlx_sam::Sam::from_safetensors_on.

pub fn compile_profile(&self) -> &CompileProfile

Tier-1 compile profile (sam.rlx.toml next to weights when present).

pub fn config(&self) -> &Sam2Config

pub fn predict_image( &mut self, image_u8: &[u8], h_in: usize, w_in: usize, points: Option<(&[f32], &[f32])>, boxes: Option<&[f32]>, mask_input: Option<&[f32]>, multimask_output: bool, ) -> Result<Sam2ImagePrediction, Error>

Image-segmentation API.

image_u8: row-major RGB h_in × w_in × 3 u8. points: optional (coords [N,2], labels [N]) — coords in input-image pixels (0..max(h_in, w_in)), labels per prompt_encoder_forward. boxes: optional [M, 4] boxes (x0, y0, x1, y1) in input pixels. mask_input: optional [1, 256, 256] low-res mask logits. multimask_output: true → 3 masks; false → 1 (with optional dynamic-stability fallback).

Returns (mask_logits, iou_pred, num_masks, h_out, w_out) where (h_out, w_out) = (4·SAM2_PROMPT_GRID, 4·SAM2_PROMPT_GRID) = 256×256 — caller resizes to the original image resolution.

pub fn predict_video_frame( &mut self, state: &mut Sam2VideoState, image_u8: &[u8], h_in: usize, w_in: usize, points: Option<(&[f32], &[f32])>, boxes: Option<&[f32]>, mask_input: Option<&[f32]>, multimask_output: bool, ) -> Result<Sam2ImagePrediction, Error>

Per-frame video API. Wraps Sam2::predict_image with the memory-attention path (cross-attend the current frame’s stride-32 features to the bank) and the memory-encoder path (encode the chosen mask + features into the bank).

Mirrors SAM2VideoPredictor.add_new_points_or_box + propagate_in_video semantics: when state is empty, this acts as image-predict; otherwise it conditions on stored frames.