Interface PreprocessedImage

interface PreprocessedImage {
    patches: Float32Array;
    positions: Int32Array;
    softTokens: number;
}

Properties

patches: Float32Array

[numSoftTokens, 6912] f32 patch vectors (padded).

positions: Int32Array

[numSoftTokens, 2] (x, y) grid positions; -1 padding.

softTokens: number

Real (unpadded) soft-token count.