DriftCapture · interface

SamPromptLayout

Where a checkpoint keeps its prompt encoder, and how its processor places a prompt: SAM's and MobileSAM's, or Transformers' SAM 2.

Explained in DriftCapture.

interface SamPromptLayout

import type { SamPromptLayout } from '@driftengine/capture';

Properties

Name	Type	Description
`gaussian`readonly	`string`	The Fourier matrix, `[2, dim / 2]`.
`embedding`readonly	`(kind: number) => readonly [name: string, row: number]`	A learned point embedding by kind — 0 off, 1 on, 2 and 3 a box's corners — as name and row.
`notAPoint`readonly	`string`
`boxPadding`readonly	`boolean`	Whether a box is followed by a padding point, as SAM 2's is.
`boxAsPoints`readonly	`boolean`	Whether a box's corners come first as points and a padding point always ends the prompt, as a SAM 2 video session orders them — where a prompt of nothing is a point meaning none, padded.
`prepared`readonly	`(height: number, width: number, size: number) => readonly [number, number]`	The prepared image's size for an original one: the longest side to `size` for SAM, the whole frame to a `size` square for SAM 2 — and whether a coordinate is scaled in double precision, as SAM's predictor does with numpy, or single, as SAM 2's processor does with a tensor.
`singleScale`readonly	`boolean`