Termux-Vision (AMEVA-Vision)
Zero-Dependency On-Device Computer Vision & Multimodal VLM Inference Engine for Android Termux
Install and configure the complete toolchain in Android Termux:
curl -sSL https://raw.githubusercontent.com/uno-km/termux-vision/main/install.sh | bashWhy Termux-Vision?
Standard computer vision and multimodal libraries (OpenCV, TorchVision, ONNX Runtime) suffer from severe compilation overhead, DPKG locks, Bionic libc symbol mismatches, and multi-gigabyte bloat on edge Android devices. Termux-Vision solves this with a native, zero-heavy-dependency architecture:
Zero-Heavy C++ Dependency
Pure Python and pure JavaScript fast paths running out-of-the-box on vanilla Termux without node-gyp or OpenCV build bottlenecks.
Multimodal VLM Inference Engine
Execute SmolVLM-500M and Qwen2-VL-2B models on-device with Vulkan GPU acceleration and automatic graceful CPU retry.
Full Classical Vision Stack
5-stage Canny edge detection, Sobel 3x3, Gaussian blur, 2D integral images, morphology, and Haar Cascade face detection.
1:1 termux-train Autograd Bridge
Extract ViT patches and vision feature maps directly into termux-train for on-device classifier and LoRA fine-tuning.
Technical Architecture
| Subsystem | Technology & Algorithmic Implementation | Execution Status |
|---|---|---|
| Classical Filters | 5-Stage Canny Edge Detector, Sobel 3x3, Gaussian Blur, 2D Integral Images | Production |
| Object & Face Detection | Haar Cascade Classifier, IoU Intersection, Non-Maximum Suppression (NMS) | Production |
| VLM Multimodal Inference | SmolVLM-500M / Qwen2-VL-2B, Supervised Process Isolation, Vulkan GPU Offload | Production |
| Neural Bridge | ViT (Vision Transformer) Patch Extraction (16x16), MobileNet Features, termux-train Bridge | Production |