Image Understanding with Vision Language Models

View Source Code

Browse the complete example on GitHub

This example demonstrates how to use Vision Language Models (VLMs) with LeapSDK on Android. VLMs combine image understanding with natural language processing, enabling your app to analyze images, answer questions about visual content, and generate detailed descriptions—all on-device.

Built with Jetpack Compose and the Coil image loading library, this example shows how to create a multimodal AI application that processes both images and text locally on Android devices.

What’s inside?

The VLMExample showcases cutting-edge multimodal AI capabilities:

• Vision Language Models - Analyze images and generate text descriptions
• Image Input Processing - Handle image selection from gallery or camera
• Multimodal Understanding - Combine visual and textual information
• Jetpack Compose UI - Modern, declarative UI for image display and results
• Coil Integration - Efficient image loading and rendering
• On-device Inference - Complete privacy with local VLM processing
• Interactive Q&A - Ask questions about images and get contextual answers

This example demonstrates the LFM2-VL-1.6B model, a vision-language model that can understand and reason about visual content.

What are Vision Language Models?

Vision Language Models (VLMs) are AI models that can process both images and text simultaneously, enabling them to:

• Describe images - Generate detailed captions of what’s in a photo
• Answer visual questions - Respond to queries about image content (“What color is the car?”)
• Detect objects - Identify and describe objects, people, and scenes
• Read text in images - Extract and interpret text from photos (OCR-like capabilities)
• Understand context - Grasp relationships between objects and spatial arrangements
• Generate insights - Provide analysis, suggestions, or interpretations of visual data

Example use cases:

• Accessibility tools that describe images for visually impaired users
• Product identification and information lookup
• Document analysis and data extraction
• Visual search and discovery
• Educational apps that explain diagrams and illustrations
• Real estate apps that describe property photos
• Medical imaging assistants for preliminary analysis

Environment setup

Before running this example, ensure you have the following:

Android Studio Installation

Download and install Android Studio (latest stable version recommended).

Make sure you have:

• Android SDK installed
• An Android device or emulator configured
• USB debugging enabled (for physical devices)

Minimum SDK Requirements

This example requires:

• Minimum SDK: API 24 (Android 7.0)
• Target SDK: API 34 or higher
• Kotlin: 1.9.0 or higher

Hardware recommendations:

• At least 4GB RAM (6GB+ recommended for better performance)
• Vision models are larger and more compute-intensive than text-only models

VLM Model Bundle Deployment

This example requires the LFM2-VL-1.6B vision language model bundle.

Step 1: Obtain the model bundle

Download the LFM2-VL-1.6B bundle from the Leap Model Library.

Step 2: Deploy to device via ADB

Use the Android Debug Bridge (ADB) to transfer the model to your device:

# Ensure your device is connected and ADB is available
adb devices

# Create the directory on device
adb shell mkdir -p /data/local/tmp/liquid/

# Push the VLM bundle to the device
adb push LFM2-VL-1_6B.bundle /data/local/tmp/liquid/

# Verify the file was transferred successfully
adb shell ls -lh /data/local/tmp/liquid/

Note: The VLM bundle is larger than text-only models (typically 1-3GB). Ensure you have sufficient storage on your device and a stable connection during transfer.

Alternative deployment location:

If /data/local/tmp/ is not accessible, use device storage:

# Push to internal storage
adb push LFM2-VL-1_6B.bundle /sdcard/Download/liquid/

# Update the model path in your app code accordingly

Dependencies Setup

Add the required dependencies to your app-level build.gradle.kts:

dependencies {
    // LeapSDK for VLM processing (0.9.7+)
    implementation("ai.liquid.leap:leap-sdk:0.9.7")

    // Coil for image loading
    implementation("io.coil-kt:coil-compose:2.5.0")

    // Jetpack Compose
    implementation(platform("androidx.compose:compose-bom:2024.01.00"))
    implementation("androidx.compose.ui:ui")
    implementation("androidx.compose.material3:material3")
    implementation("androidx.compose.ui:ui-tooling-preview")
    implementation("androidx.activity:activity-compose:1.8.2")

    // Image picker
    implementation("androidx.activity:activity-ktx:1.8.2")

    // ViewModel
    implementation("androidx.lifecycle:lifecycle-viewmodel-compose:2.7.0")
}

About Coil:

Coil is a Kotlin-first image loading library for Android that:

• Efficiently loads and caches images
• Integrates seamlessly with Jetpack Compose
• Handles image transformations and processing
• Provides modern coroutine-based APIs

How to run it

Follow these steps to start analyzing images with VLMs:

1. Clone the repository

   git clone https://github.com/Liquid4All/LeapSDK-Examples.git
   cd LeapSDK-Examples/Android/VLMExample
   

2. Deploy the VLM model bundle

   • Follow the ADB commands in the setup section above
   • Ensure the bundle is accessible at /data/local/tmp/liquid/LFM2-VL-1_6B.bundle

3. Open in Android Studio

   • Launch Android Studio
   • Select “Open an existing project”
   • Navigate to the VLMExample folder and open it

4. Verify model path

   • Check that the model path in your code matches the deployment location
   • Update if you used a different path

5. Run the app

   • Connect your Android device or start an emulator
   • Click “Run” or press Shift + F10
   • Select your target device

6. Select an image

   • On first launch, the app will load the VLM model (this may take 10-30 seconds)
   • Tap the “Select Image” button
   • Choose an image from your device’s gallery
   • Alternatively, take a photo if camera integration is enabled

7. Analyze the image

   • After selecting an image, it will be displayed in the app
   • The VLM will automatically analyze the image
   • View the generated description or ask questions about the image
   • Try different prompts: “What’s in this image?”, “Describe the scene”, “What colors do you see?”

Performance Note: Vision models are computationally intensive. First-time inference may take 5-15 seconds on mobile devices. Subsequent inferences on the same or similar images will be faster as the model stays loaded in memory.

Understanding the architecture

Image Selection Flow

The app uses Android’s image picker to select photos:

@Composable
fun VLMScreen(viewModel: VLMViewModel) {
    val imagePickerLauncher = rememberLauncherForActivityResult(
        contract = ActivityResultContracts.GetContent()
    ) { uri: Uri? ->
        uri?.let { viewModel.processImage(it) }
    }

    Button(onClick = { imagePickerLauncher.launch("image/*") }) {
        Text("Select Image")
    }
}

VLM Integration Pattern

Loading and using the vision language model:

class VLMViewModel : ViewModel() {
    private lateinit var vlmModel: LeapVLModel

    fun initializeModel() {
        viewModelScope.launch(Dispatchers.Default) {
            vlmModel = LeapSDK.loadVLModel(
                bundlePath = "/data/local/tmp/liquid/LFM2-VL-1_6B.bundle"
            )
            _modelState.value = ModelState.Ready
        }
    }

    fun processImage(imageUri: Uri) {
        viewModelScope.launch(Dispatchers.Default) {
            // Load image from URI
            val bitmap = loadBitmapFromUri(imageUri)

            // Generate description
            val prompt = "Describe this image in detail."
            val description = vlmModel.generateFromImage(
                image = bitmap,
                prompt = prompt,
                maxTokens = 200
            )

            _imageAnalysis.value = ImageAnalysis(
                imageUri = imageUri,
                description = description
            )
        }
    }

    private fun loadBitmapFromUri(uri: Uri): Bitmap {
        return context.contentResolver.openInputStream(uri)?.use { inputStream ->
            BitmapFactory.decodeStream(inputStream)
        } ?: throw IllegalArgumentException("Unable to load image")
    }
}

Coil Integration for Image Display

Using Coil to efficiently display selected images:

@Composable
fun ImageAnalysisDisplay(analysis: ImageAnalysis) {
    Column(
        modifier = Modifier
            .fillMaxSize()
            .padding(16.dp)
    ) {
        // Display image with Coil
        AsyncImage(
            model = ImageRequest.Builder(LocalContext.current)
                .data(analysis.imageUri)
                .crossfade(true)
                .build(),
            contentDescription = "Selected image",
            modifier = Modifier
                .fillMaxWidth()
                .height(300.dp)
                .clip(RoundedCornerShape(8.dp)),
            contentScale = ContentScale.Crop
        )

        Spacer(modifier = Modifier.height(16.dp))

        // Display AI-generated description
        Card(
            modifier = Modifier.fillMaxWidth()
        ) {
            Column(modifier = Modifier.padding(16.dp)) {
                Text(
                    text = "Analysis",
                    style = MaterialTheme.typography.titleMedium
                )
                Spacer(modifier = Modifier.height(8.dp))
                Text(
                    text = analysis.description,
                    style = MaterialTheme.typography.bodyMedium
                )
            }
        }
    }
}

Interactive Q&A Mode

Allow users to ask questions about images:

fun askQuestionAboutImage(bitmap: Bitmap, question: String): String {
    return vlmModel.generateFromImage(
        image = bitmap,
        prompt = "Answer this question about the image: $question",
        maxTokens = 150
    )
}

// Example usage
val answer1 = askQuestionAboutImage(bitmap, "What is the main object in this image?")
val answer2 = askQuestionAboutImage(bitmap, "What colors are prominent?")
val answer3 = askQuestionAboutImage(bitmap, "Is this indoors or outdoors?")

Memory Management

Vision models require more memory. Implement proper lifecycle handling:

override fun onStop() {
    super.onStop()
    // Release model when app goes to background to free memory
    viewModel.releaseModel()
}

override fun onStart() {
    super.onStart()
    // Reload model when app returns to foreground
    viewModel.initializeModel()
}

Results

The VLMExample demonstrates powerful image understanding capabilities:

The interface shows:

• Selected image displayed clearly with Coil
• AI-generated analysis below the image
• Smooth, responsive UI even with large images
• Professional Material3 design

Example analysis output:

Image: A sunset over a beach

"The image shows a beautiful sunset scene at a beach. The sky displays
vibrant orange and pink hues as the sun sets on the horizon. The ocean
water reflects the warm colors of the sky. In the foreground, there are
silhouettes of people walking along the shoreline. The overall mood is
peaceful and serene."

All processing happens entirely on your Android device, ensuring complete privacy for your photos.

Further improvements

Here are some ways to extend this example:

• Camera integration - Take photos directly in-app for immediate analysis
• Multi-image support - Compare and analyze multiple images simultaneously
• Batch processing - Process entire photo albums with progress tracking
• Custom prompts - Let users enter their own questions about images
• Object detection - Highlight detected objects with bounding boxes
• Text extraction - Pull out text from images (receipts, documents, signs)
• Image editing suggestions - Recommend crops, filters, or enhancements
• Accessibility features - Auto-generate alt text for images
• Favorites and history - Save analyzed images with their descriptions
• Export functionality - Share analysis results or create reports
• Comparison mode - Analyze differences between two images
• Real-time video analysis - Process camera frames in real-time
• Multilingual descriptions - Generate descriptions in different languages
• Style transfer guidance - Describe artistic styles and suggest transformations

Need help?

Join our Discord

Connect with the community and ask questions about this example.