Create the Android Studio LiteRT project

java -version
adb version
# In Android Studio: New Project → Empty Activity → Kotlin → min SDK 23.
# Copy the converted model into app/src/main/assets/.
mkdir -p app/src/main/assets
cp clothing.tflite app/src/main/assets/clothing.tflite

Gradle dependency

// app/build.gradle.kts
android {
    namespace = "com.example.clothinglitert"
    compileSdk = 35
    defaultConfig {
        applicationId = "com.example.clothinglitert"
        minSdk = 23
        targetSdk = 35
        versionCode = 1
        versionName = "1.0"
    }
}
dependencies {
    // Published CompiledModel artifact used by this lesson.
    implementation("com.google.ai.edge.litert:litert:2.1.5")
}

Sync Gradle. LiteRT 2.x gives you the current CompiledModel API. You pick CPU, GPU, or NPU through options. This course starts with CPU so the first results are easy to repeat. The older Interpreter API is not used for new code.

Package and load the asset

Files in app/src/main/assets live inside the APK. They are not ordinary file paths on disk. Copy the asset to the app’s files directory once, then pass that absolute path to the API.

package com.example.clothinglitert

import android.content.Context
import java.io.File

fun copyModelFromAssets(context: Context): File {
    val destination = File(context.filesDir, "clothing.tflite")
    if (!destination.exists()) {
        context.assets.open("clothing.tflite").use { input ->
            destination.outputStream().use { output -> input.copyTo(output) }
        }
    }
    check(destination.length() > 0) { "clothing.tflite was not packaged" }
    return destination
}

Create the compiled model and buffers

import com.google.ai.edge.litert.Accelerator
import com.google.ai.edge.litert.CompiledModel

val modelFile = copyModelFromAssets(this)
val compiledModel = CompiledModel.create(
    modelFile.absolutePath,
    CompiledModel.Options(Accelerator.CPU),
)
val inputBuffers = compiledModel.createInputBuffers()
val outputBuffers = compiledModel.createOutputBuffers()
check(inputBuffers.size == 1) { "expected one input buffer" }
check(outputBuffers.size == 1) { "expected one output buffer" }

Create the compiled model and its buffers once, for example in a ViewModel or when the activity starts. Do not create them for every selected image. The path-based overload is the documented quick start. If your pinned SDK has a verified asset-manager overload, you can use that instead.

First APK check

1. Build and install the empty app.
2. Open APK Analyzer and confirm assets/clothing.tflite is present.
3. Run the asset-copy check and log its non-zero size.
4. Only then add image UI and inference.