ArAnirudh/slimsam-77-uniform-boxes

SlimSAM-77 (uniform) for transformers.js — with box prompts

SlimSAM (nielsr/slimsam-77-uniform) in ONNX for transformers.js, like Xenova/slimsam-77-uniform — whose image encoder files are reused here unchanged — but with a prompt encoder / mask decoder that takes box prompts.

Xenova/slimsam-77-uniform's decoder declares no input_boxes, and transformers.js drops inputs a session does not declare, so a box passed to it is silently ignored. This decoder follows Meta's own SAM ONNX convention instead: a box is two extra points in input_points, labelled 2 (top-left corner) and 3 (bottom-right corner). Inputs and outputs keep transformers.js's names, so SamModel / SamProcessor drive it unchanged.

One rule the caller must follow: a prompt without a box carries a padding point (0, 0) with label -1 (what Hugging Face's SamPromptEncoder adds internally when no box is given). A prompt with a box has no pad.

// click: one positive point + the pad
input_points: [[[[x, y], [0, 0]]]], input_labels: [[[1, -1]]]
// box with a seed point: point, then the two corners
input_points: [[[[x, y], [x0, y0], [x1, y1]]]], input_labels: [[[1, 2, 3]]]

file
onnx/prompt_encoder_mask_decoder.onnx	fp32 (WebGPU default) — matches PyTorch SamModel exactly on points, points + box
onnx/prompt_encoder_mask_decoder_quantized.onnx	dynamic uint8 (WASM default) — 99.2–99.96% mask agreement with fp32
onnx/vision_encoder*.onnx, config.json, preprocessor_config.json	from Xenova/slimsam-77-uniform, unchanged

Exported by export_decoder.py (in this repo), which also checks the export against PyTorch. Apache-2.0, as the SlimSAM weights and the Xenova export it builds on.