SlimSAM-77 uniform segment anything (int8)
Image segmentation model for onnxruntime-web (int8 dynamic-quantized). Runs on WASM —
13.1 MB downloaded once from models.skillsafe.ai, then cached for every SkillSafe app that
uses it. Inference happens on your device; nothing you enter is uploaded to load it, and it never costs a credit.
Weights from Hugging Face · Xenova/slimsam-77-uniform, pinned at 5850ab45f587 — also loadable straight from Hugging Face outside SkillSafe (how).
Details
slimsam-77-uniform-int8@5850ab455850ab45f587c112167512ffef949107115e26a0Files
Each file is served at an immutable URL; the canonical form is the SHA-256 itself. Tokenizer and config JSON are not here by design — they ship in your app bundle.
| Path | Format | Size | SHA-256 | |
|---|---|---|---|---|
onnx/vision_encoder_quantized.onnx | onnx | 8.5 MB | cce23c7b2e5d…1f2971 | |
onnx/prompt_encoder_mask_decoder_quantized.onnx | onnx | 4.7 MB | cb90b279f549…7fdce2 |
Signature
Graph inputs and outputs read from the ONNX bytes at vetting — the tensor names your session.run() call feeds and reads. Symbolic dimensions are shown by name.
onnx/vision_encoder_quantized.onnx
Inputs
pixel_valuesfloat32 [batch_size, 3, 1024, 1024]
Outputs
image_embeddingsfloat32 [batch_size, 256, 64, 64]image_positional_embeddingsfloat32 [batch_size, 256, 64, 64]
onnx/prompt_encoder_mask_decoder_quantized.onnx
Inputs
input_pointsfloat32 [batch_size, point_batch_size, nb_points_per_image, 2]input_labelsint64 [batch_size, point_batch_size, nb_points_per_image]image_embeddingsfloat32 [batch_size, 256, 64, 64]image_positional_embeddingsfloat32 [batch_size, 256, 64, 64]
Outputs
iou_scoresfloat32 [batch_size, point_batch_size, 3]pred_masksfloat32 [batch_size, point_batch_size, 3, 256, 256]
Use it in an app declaration · SDK loader · onnxruntime-web · URLs · Hugging Face — generated from this entry
Add to the body of POST /v1/apps/{slug}/releases (or a release session). An unknown or withdrawn model is refused with a 400 naming it; the app page then shows "downloads 13.1 MB · runs on your device" and /models.txt carries the attribution.
{
"models": [
{
"id": "slimsam-77-uniform-int8",
"revision": "5850ab45"
}
]
}Streams with progress, verifies the SHA-256 against the catalogue, keeps a durable copy in the app's Cache API and reports whether the bytes came from cache or the network.
<script src="/sdk.js"></script> <!-- vendored from https://skillsafe.ai/apps-sdk/v1.js -->
const ss = SkillSafe.init({ slug: "your-app" });
const caps = await ss.models.capabilities(); // { webgpu, wasm, wasmSimd, storage }
const files = await ss.models.loadAll("slimsam-77-uniform-int8@5850ab45", {
onProgress: (p) => (bar.style.width = Math.round(p.ratio * 100) + "%"),
}); // { path: ArrayBuffer } — SHA-256 verified, cached
const vision_encoder_quantized = files["onnx/vision_encoder_quantized.onnx"];
const prompt_encoder_mask_decoder_quantized = files["onnx/prompt_encoder_mask_decoder_quantized.onnx"];
// Later visits: await ss.models.status("slimsam-77-uniform-int8@5850ab45") → { cached: [...], missing: [] }
// Background prefetch on a landing page: ss.models.warm("slimsam-77-uniform-int8@5850ab45")The runtime itself ships in your bundle as a vetted {path, sha256} reference — only the weights come from the registry. Pass the bytes in; do not re-fetch by URL.
// onnxruntime-web 1.17.0 ships in your bundle as a vetted {path, sha256} reference —
// dist/ort.wasm.bundle.min.mjs + dist/ort-wasm-simd-threaded.wasm.
import * as ort from "./ort.wasm.bundle.min.mjs";
ort.env.wasm.wasmPaths = { wasm: "/ort-wasm-simd-threaded.wasm" }; // object form, never a prefix string
ort.env.wasm.numThreads = 1; // app hosts have no COOP/COEP
const bytes = await ss.models.load("slimsam-77-uniform-int8@5850ab45", "onnx/vision_encoder_quantized.onnx", { onProgress });
const session = await ort.InferenceSession.create(new Uint8Array(bytes), {
executionProviders: ["wasm"],
});
// Feeds are named after the graph inputs (read from the file, see the signature above):
const feeds = {
"pixel_values": new ort.Tensor("float32", new Float32Array(n), [/* batch_size */ 1, 3, 1024, 1024]),
};
const out = await session.run(feeds);
const image_embeddings = out["image_embeddings"].data; // float32 [batch_size, 256, 64, 64]
const image_positional_embeddings = out["image_positional_embeddings"].data; // float32 [batch_size, 256, 64, 64]Immutable, credential-free, Access-Control-Allow-Origin: *, range requests honoured. The host refuses requests carrying a foreign Origin; a plain curl is fine.
https://models.skillsafe.ai/slimsam-77-uniform-int8@5850ab45/onnx/vision_encoder_quantized.onnx
https://models.skillsafe.ai/slimsam-77-uniform-int8@5850ab45/onnx/prompt_encoder_mask_decoder_quantized.onnxcurl -sSL -o vision_encoder_quantized.onnx "https://models.skillsafe.ai/slimsam-77-uniform-int8@5850ab45/onnx/vision_encoder_quantized.onnx"
shasum -a 256 vision_encoder_quantized.onnx # cce23c7b2e5d4f330932738fb67ba518e04b0d99ccdd1cccd22a7da4e01f2971The Hugging Face repo these bytes were vetted from, pinned at 5850ab45f587. Use it directly for your own site or notebook. A SkillSafe app must load from the registry: its CSP allows models.skillsafe.ai only, and that copy is what the SDK verifies, the browser shares across apps and the app page discloses. SkillSafe keeps hosting its own copies and the API either way.
// Outside SkillSafe apps only: app hosts allow connect-src models.skillsafe.ai, not huggingface.co.
// Same bytes as the registry copy — the SHA-256 on the Hugging Face file page is cce23c7b2e5d4f33…
const res = await fetch("https://huggingface.co/Xenova/slimsam-77-uniform/resolve/5850ab45f587c112167512ffef949107115e26a0/onnx/vision_encoder_quantized.onnx");
const bytes = new Uint8Array(await res.arrayBuffer());
const session = await ort.InferenceSession.create(bytes, { executionProviders: ["wasm"] });https://huggingface.co/Xenova/slimsam-77-uniform/resolve/5850ab45f587c112167512ffef949107115e26a0/onnx/vision_encoder_quantized.onnx
https://huggingface.co/Xenova/slimsam-77-uniform/resolve/5850ab45f587c112167512ffef949107115e26a0/onnx/prompt_encoder_mask_decoder_quantized.onnxEvaluation
onnx.checker + onnxruntime CPU smoke run with zero-filled inputs at the declared shapes; per-file SHA-256 pinned to the source; ONNX output vs the PyTorch model from the same commit on real text (import.parity) — imported as published upstream, then checked. Evaluated Sep 22, 2026.
Parity against transformers SamVisionModel from nielsr/slimsam-77-uniform@72a5ae6ca4d1, fp32 · last_hidden_state
| file | precision | max abs | mean abs | cosine / PSNR |
|---|---|---|---|---|
onnx/vision_encoder_quantized.onnx | quantised | 0.059 | 6.1e-3 | 0.998347 |
Runs under onnxruntime
Zero-filled inputs at the declared shapes, CPU execution provider on the converter host; the check is that the graph loads, runs, and emits the declared output shapes.
onnx/prompt_encoder_mask_decoder_quantized.onnx: input_points[1,1,1,2], input_labels[1,1,1], image_embeddings[1,256,64,64], image_positional_embeddings[1,256,64,64] → iou_scores[1,1,3], pred_masks[1,1,3,256,256] 11.9 msonnx/vision_encoder_quantized.onnx: pixel_values[1,3,1024,1024] → image_embeddings[1,256,64,64], image_positional_embeddings[1,256,64,64] 388.4 ms
Toolchain: python 3.12.13 · platform Darwin 25.6.0 arm64 · torch 2.10.0 · onnx 1.23.0 · onnxruntime 1.30.0. Recipe models/recipes/slimsam-77-uniform-int8.yaml (33c47dabbe64). Full manifest.json
Licence & attribution
Apache-2.0 · licence text · notice
SlimSAM: 0.1% Data Makes Segment Anything Slim (Chen et al., NeurIPS 2024), a structural pruning of Meta's Segment Anything ViT-B. ONNX int8 dynamic-quantized export of the 77%-pruned uniform model published by Xenova on Hugging Face. Licensed under the Apache License 2.0.
Apps that declare this model get this text in their generated /models.txt, so a licence that requires a notice always carries one.
Every file here was approved by exact SHA-256 after a structural audit of the graph, fetched from a content-pinned source, and is served credential-free at an immutable URL. The SDK re-verifies the hash on your device before it caches or returns anything. Missing a variant? Request it — or read how the registry works.