OpenCV 5 Ships New DNN Engine with 80% ONNX Coverage

OpenCV 5 is here with a completely rewritten DNN engine that boosts ONNX operator support from 22% to 80%, adds native LLM/VLM inference, and ships hardware acceleration improvements. The new graph-based engine fuses transformer attention ops, runs models like YOLOv8n 11.5% faster than ONNX Runtime, and keeps the classic engine for backward compatibility.

2 min readJun 9, 2026

OpenCV 5 Ships New DNN Engine with 80% ONNX Coverage

OpenCV 5 drops June 8th (pip version) with a DNN engine rewrite that jumps ONNX operator coverage from ~22% to over 80%. The new graph-based engine understands dynamic shapes, fuses attention patterns, and runs LLMs like Qwen 2.5 directly inside the library.

Three Engines, One API

OpenCV 5 keeps the old engine for backward compatibility. You pick at load time via the EngineType enum:

import cv2 as cv
# Default: new engine first, fallback to classic
net = cv.dnn.readNetFromONNX(&#34;model.onnx&#34;)
# Force new engine
net = cv.dnn.readNetFromONNX(&#34;model.onnx&#34;, engine=cv.dnn.ENGINE_NEW)

The classic engine (ENGINE_CLASSIC) supports non-CPU backends like CUDA and OpenVINO. The new engine (ENGINE_NEW) runs on CPU only for now but delivers fusions like FlashAttention-style attention collapsing. ENGINE_ORT wraps ONNX Runtime if built with WITH_ONNXRUNTIME=ON.

Speed Benchmarks

On an Intel Core i9-14900KS (Ubuntu 24.04), OpenCV 5's new engine beats ONNX Runtime on several models:

Model	OpenCV 5 DNN (ms)	ONNX Runtime (ms)	Difference
XFeat	6.56	8.61	31.25% faster
YOLOv8n	10.9	12.15	11.5% faster
YOLOX-S	23.46	25.16	7.24% faster
DINOv2 small	23.78	29.58	24.4% faster
RF-DETR	102.01	106.49	4.4% faster
OWLv2	1,090	1,489	36.6% faster
BiRefNet	7,178	9,503	32.4% faster

LLMs and VLMs Inside OpenCV

OpenCV 5 ships a native tokenizer and KV-cache for autoregressive decoding. Models like Qwen 2.5, Gemma 3, PaliGemma, and GPT-2/4 run through the same Net API as YOLO. In tests, Qwen 2.5 output matched ONNX Runtime token-for-token.

Other Improvements

Core modernization: Retired legacy C API, added native FP16/BF16, proper 0D/1D tensors, and real logging.
Hardware acceleration: Cleaner layer for vendor-specific kernels.
3D vision: ChArUco boards, multi-camera calibration, and better visualization.
Documentation: Completely revamped, easier to navigate.
Models validated: YOLOv8, YOLOv9, YOLOv10, DINOv2, SAM, CLIP, RT-DETR, LaMa inpainting, and more.

What's Next

GPU support in the new DNN engine and a non-CPU HAL for accelerated pre/post-processing are planned. The OpenCV team invites community testing and contributions.

Get the pip version on June 8th. For now, build from the GitHub master branch.

Editor's Take

I've been burned by OpenCV's DNN module refusing to load modern ONNX models more times than I can count. The jump from 22% to 80% operator coverage is huge — it means I can finally ditch the ONNX Runtime wrapper in my production pipelines. That said, the new engine being CPU-only is a bummer; I'll still need the classic engine for GPU inference. I'm most excited about the LaMa inpainting demo — removing objects without an extra framework is exactly the kind of integration that makes OpenCV indispensable.

— DevDigest Editorial

Key Takeaways

•Use ENGINE_AUTO (default) for zero-config upgrades; fallback to classic engine ensures existing models still work.
•Pin ENGINE_NEW for CPU inference on transformer models to get automatic attention fusion and FlashAttention.
•For GPU inference, stick with ENGINE_CLASSIC and setPreferableBackend(CUDA) — new engine is CPU-only for now.

Why It Matters

OpenCV 5 eliminates the need for a separate ONNX Runtime or LLM serving stack for many vision pipelines. With 80% ONNX coverage and native LLM inference, you can now run detection, segmentation, and captioning in a single dependency. The upgrade risk is low thanks to the multi-engine fallback.

#machine-learning#open-source#ONNX#computer vision#OpenCV

Get the weekly digest

Every Sunday - top tech stories, industry breakthroughs, and developer tools delivered to your inbox.

No spam, unsubscribe anytime.