Termux-Vision

Quickstart & Recipes

Practical minimal code recipes for Python, Node.js, and CLI workflows

Recipe 1: Multimodal VLM Image Chat (Python)

import termux_vision as tv

# Load VLM Engine (Auto Vulkan GPU with CPU fallback)
with tv.vlm.load(model_id="smolvlm-500m-q4", device="auto") as engine:
    result = engine.describe(
        "photo.jpg",
        prompt="이 사진 속 인물과 배경을 요약 설명해줘.",
        max_tokens=150
    )
    print(f"[Result]: {result.text}")
    print(f"[Speed]: {result.metrics.tokens_per_second:.1f} t/s")

Recipe 2: Multimodal VLM Image Chat (Node.js)

const tv = require('termux-vision');

async function main() {
  const engine = await tv.vlm.load({
    modelId: 'smolvlm-500m-q4',
    device: 'auto'
  });

  const res = await engine.describe('photo.jpg', {
    prompt: 'Explain what is happening in this image.'
  });

  console.log(`[Result]: ${res.text}`);
  console.log(`[Decode]: ${res.metrics.decodeMs} ms`);
}
main();

Recipe 3: Classical Canny Edge Detection (Python)

import termux_vision as tv

# Load, resize, and convert to grayscale
img = tv.io.load_image("input.jpg")
gray = tv.transforms.to_grayscale(tv.transforms.resize(img, (512, 512)))

# Execute 5-stage Canny Edge Detection
edges = tv.cv.canny(gray, low_threshold=40.0, high_threshold=120.0)
tv.io.save_image(edges, "edges.png")
print("Edges saved to edges.png")

Recipe 4: Haar Cascade Face Detection

import termux_vision as tv

img = tv.io.load_image("selfie.jpg")
detections = tv.detect.detect_faces(img)

print(f"Found {len(detections)} candidate face regions:")
for d in detections:
    print(f"  - BBox: {d.bbox.to_xywh()} (Score: {d.score})")