yolo26-kit

Drop YOLO26 into your existing YOLOv8 pipeline. Bridge library for the NMS-free, end-to-end era.

Why

YOLO26 default ONNX export is end2end=True, producing (N, 300, 6) already-decoded detections. That's great — but it breaks every Triton / DeepStream / OpenCV DNN config that expects YOLOv8-style (1, 4+nc, N). It also leaves you to write letterbox-coord undo math by hand. And some targets (RKNN, certain TFLite quant paths) can't fuse the e2e head, so they ship raw (1, 4+nc, N) you have to decode yourself.

yolo26-kit is a small, pure-functional library that bridges that gap — and nothing else.

Install

# Python — https://pypi.org/project/yolo26-kit/
pip install yolo26-kit            # core
pip install "yolo26-kit[ort]"     # + ORT wrapper

# TypeScript / JavaScript — https://www.npmjs.com/package/yolo26-kit
npm i yolo26-kit                 # core
npm i yolo26-kit onnxruntime-web # + ORT wrapper (peer dep)

Export YOLO26 to ONNX

The "happy path" is to export with end2end=True so the ONNX graph emits already-deduplicated (N, 300, 6) detections:

from ultralytics import YOLO

YOLO("yolo26n.pt").export(format="onnx", end2end=True)
# produces yolo26n.onnx with output shape (1, 300, 6)

If you cannot use end2end=True (e.g., quantized TFLite or some NPU runtimes that don't support the e2e head), yolo26-kit auto-detects and runs class-aware NMS for you when decoding raw (1, 4+nc, N) outputs.

Use

Python — e2e default (most users)

import onnxruntime as ort
import yolo26_kit

sess = ort.InferenceSession("yolo26n.onnx")
out = sess.run(None, {"images": pre})[0]   # (1, 300, 6)
out = yolo26_kit.normalize_output(out)
dets = yolo26_kit.filter_e2e(out, conf=0.25)
# [{'box': [...], 'score': ..., 'class': ..., 'label': '...'}]

Python — drop into existing YOLOv8 pipeline

out_v8 = yolo26_kit.e2e_to_v8_shape(out)   # (1, 84, 300)
# Feed out_v8 to your existing v8 Triton / DeepStream / OpenCV DNN parser.

Python — convenience wrapper (image in, dets in original coords out)

from yolo26_kit import from_ort

decoder = from_ort("yolo26n.onnx")
dets = decoder.predict("bus.jpg", conf=0.25)

TypeScript

import { filterE2E } from "yolo26-kit";
import { fromOrt } from "yolo26-kit/ort";
import * as ort from "onnxruntime-web";

const session = await ort.InferenceSession.create("yolo26n.onnx");
const decoder = fromOrt(session);
const dets = await decoder.predict(canvas, { conf: 0.25 });

Single image at a time (batch=1). Loop in your code if you have multiple frames.

TypeScript — non-e2e raw export

import { decodeDetect, normalizeOutput } from "yolo26-kit";

const out = normalizeOutput(rawTensor);
const dets = decodeDetect(out, [1, 84, 8400], { conf: 0.25, numClasses: 80 });

Decoder auto-routing

The Decoder auto-detects whether the ONNX is e2e (output shape ends in 6) or raw (output shape (1, 4+nc, N)):

• e2e: runs filter_e2e — already deduped by the model, no NMS.
• raw: runs decode_detect with class-aware NMS (default iou_threshold=0.45).

Override the NMS default if you want raw outputs (e.g., for further post-processing):

decoder.predict("bus.jpg", conf=0.25, nms=False)

await decoder.predict(canvas, { conf: 0.25, nms: false });

What's in the box (v0.1.x)

Compatibility

A nightly live-diff workflow installs the latest ultralytics and re-validates fixtures. Drift opens an auto-tagged drift issue.

Status

v1 (this release): Detect task only. e2e + non-e2e paths. Both Python and TypeScript, cross-language equivalent (verified via golden fixtures).

Roadmap: Seg, pose, cls, OBB tasks. Reproduction kit (Docker + COCO train/eval). ONNX schema standardizer. NPU ports (RKNN, Hailo, RDK, Qualcomm SNPE/QNN).

Development

# Python
cd python
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"
pytest tests/ -v
mypy src/yolo26_kit
ruff check src tests

# TypeScript
cd js
npm install
npm test
npm run typecheck
npm run lint

Limitations (v0.1.x)

• Single image only: decoders reject batch > 1. Loop in user code for multi-image inference. Batched decode tracked for v0.2.
• Detect task only: seg/pose/cls/OBB queued.
• Browser preprocess uses nearest-neighbor resize for portability. For pixel-perfect parity with PIL bilinear, pre-resize via canvas.

License

Apache-2.0 for the code in this repository. Model weights are not redistributed; users fetch them from the official Ultralytics distribution under its own terms (AGPL-3.0).

See fixtures/LICENSES.md for fixture asset attribution.