visionai @visionaidev

FastVLM - low latency VLM models. ONNX model can be seen in hugging face. 7B was good, but I think 0.5B size is maximum for practical use on edge computers without GPU. I'll try to implement to ezonnx library. github.com/apple/ml-fastv…

過去に試した様々な画像AIモデルの推論コードを整理してGitHubで公開しました。基本2行でモデルロードから推論まで完了、と簡単に使えます。ONNXだからtorch環境にも依存なし。まだモデル追加していくので是非チェックしてみて下さい github.com/ikeboo/ezonnx

ALIKED-LightGlue - image matching method with local feature descriptor ALIKED and keypoint matcher LightGlue. Now it has come to ezonnx library. No complex process and environment, just add 2 lines of python code to get result. github.com/ikeboo/ezonnx #computervision #ai #python

PPOCRv5 - Fast, light weight but good accuracy OCR model. It came to ezonnx library! Super easy to use, no dependency on torch, tf, pp. dots.ocr or MiniCPM-V 4.5 are still heavy for ocr on edge. Check ppocrv5 in ezonnx! github.com/ikeboo/ezonnx

I caught a gecko, wait... it has two tails! 壁についてたヤモリ捕まえたら、尻尾が2本あるぞ、、ヒトカゲみたい #Charmander

MotionBERT 3D arrived at ezonnx - 3D pose estimation model. You can directly get 3D poses from the video with just 2 lines of python code. Please visit ezonnx repository. github.com/ikeboo/ezonnx

keyboard_arrow_left Previous keyboard_arrow_right Next

RTMW - Person whole body pose detection model. State-of-the-art level accuracy including detail of face and hand. Easy to use with ezonnx. Write Just 2 lines of python code, and you can get awesome detection result. github.com/ikeboo/ezonnx

RTMPose - person pose detection model. 17 points and 133 points models are available. Super easy to use with ezonnx. Check GitHub repository github.com/ikeboo/ezonnx #ai #machinelearing #Python

Real-ESRGAN - image super resolution model. Not latest, but enough quality for practical use. Added to ezonnx library, you can use it just with 2 lines of code. github.com/ikeboo/ezonnx

What is standard model in super resolution these days? SuperESRGAN is good, but old now. Recent models are difficult because of their license and size(diffusion model is too heavy for the practical use)

I added optical flow estimation model NeuFlowV2 to ezonnx. And over 10 variations of vision tasks can be handled with simple and unified code. Still I have past works, I will add image matching, super resolution and ocr soon. github.com/ikeboo/ezonnx

I created GitHub repo to handle SOTA vision ONNX models easily, including DINOv3, SigLIP2, YOLO, DWPose, SAM2, D-FINE, DepthAnythingV2 and others. Simple and unified API. It’s thanks to the amazing models from the Hugging Face onnx-community. github.com/ikeboo/ezonnx

Tested one-shot anomaly detection with DINOv3 ViTsplus. Just calculate similarity of patch tokens of an anomaly image to the single good image. As I expected the anomaly area was detected. Good. Anomaly detection method with DINOv2 (anomalyDINO and dinomaly) will be improved.

Testing visual prompt object detection and segmentation only with DINOv3. Left is prompt by mouseover and right is target image.

DINOv3 , using its high quality embeddings, it works like the visual prompt segmentation. In my demo, prompt by mouseover on left image, segmentation on right. Great!

I tested zero-shot or few-shot capability of DINOv3. Surprisingly the selected area in reference image(left) is correctly highlighted in target image(right) even target is not common image. This must be game changer in the object detection by visual prompt.

DINOv3 - Image feature extraction model. I tried feature map extraction and got much better feature map clearly. SOTA models in many different tasks depend on DINOv2, they will improve its scores with DINOv3! github.com/facebookresear…

anomalib - image anomaly detection library for python. Good news, the version 2.1 contains "dinomaly" and "UniNet". These SOTA methods require to update parameter on training but their scores on the test dataset seems good. github.com/open-edge-plat…

Qwen3-4B-Instruct-2507 : Great update, and the score improvement seems like it would be worth a major update. huggingface.co/Qwen/Qwen3-4B-…

kitten-tts-nano, text to speech model with small parameter. Parameter size is just 20% compared to kokoro tts. I felt the voice quality is little bit worse than kokoro, but it's enough for the kind of voice chat or practical communications. huggingface.co/KittenML/kitte…