<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Tushar ahmed</title>
    <description>The latest articles on DEV Community by Tushar ahmed (@tusharahmed11).</description>
    <link>https://dev.to/tusharahmed11</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F985929%2Fb509fa82-b039-44f3-9d06-d97c09a7cb75.jpeg</url>
      <title>DEV Community: Tushar ahmed</title>
      <link>https://dev.to/tusharahmed11</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/tusharahmed11"/>
    <language>en</language>
    <item>
      <title>Building an Offline Food Scanner on Android with Google AI Edge LiteRT and CameraX</title>
      <dc:creator>Tushar ahmed</dc:creator>
      <pubDate>Mon, 16 Nov 2026 16:33:32 +0000</pubDate>
      <link>https://dev.to/tusharahmed11/building-an-offline-food-scanner-on-android-with-google-ai-edge-litert-and-camerax-1gfc</link>
      <guid>https://dev.to/tusharahmed11/building-an-offline-food-scanner-on-android-with-google-ai-edge-litert-and-camerax-1gfc</guid>
      <description>&lt;p&gt;Running machine learning models completely offline on Android gives you zero latency, zero API costs, and full user privacy. Google's &lt;strong&gt;LiteRT&lt;/strong&gt; (&lt;code&gt;com.google.ai.edge.litert&lt;/code&gt;, the evolution of TensorFlow Lite) makes it straightforward to run &lt;code&gt;.tflite&lt;/code&gt; models directly on-device.&lt;/p&gt;

&lt;p&gt;In this tutorial, we'll walk through how &lt;strong&gt;NutriLens&lt;/strong&gt; — a Jetpack Compose + CameraX food &amp;amp; nutrition tracker — uses LiteRT to identify dishes from a live camera feed and estimate calories, macros, and micronutrients &lt;strong&gt;100% offline&lt;/strong&gt;.&lt;/p&gt;




&lt;h2&gt;
  
  
  What We Are Building
&lt;/h2&gt;

&lt;p&gt;The NutriLens scanning pipeline works like this:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;CameraX ImageAnalysis (RGBA_8888)
        │
        ▼
Rotate upright &amp;amp; throttle (~850ms)
        │
        ▼
Center-crop &amp;amp; scale Bitmap (192×192)
        │
        ▼
LiteRT Inference (CompiledModel / Interpreter)
        │
        ▼
Dequantize UINT8 Softmax → Top-3 Food Labels
        │
        ▼
Temporal Smoothing + Nutrition Estimator → Compose UI
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;






&lt;h2&gt;
  
  
  Step 1: Add LiteRT and CameraX Dependencies
&lt;/h2&gt;

&lt;p&gt;In &lt;code&gt;gradle/libs.versions.toml&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight toml"&gt;&lt;code&gt;&lt;span class="k"&gt;[versions]&lt;/span&gt;
&lt;span class="na"&gt;camerax&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;"1.4.2"&lt;/span&gt;
&lt;span class="na"&gt;litert&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;"2.1.0"&lt;/span&gt;

&lt;span class="k"&gt;[libraries]&lt;/span&gt;
&lt;span class="na"&gt;androidx-camera-core&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;group&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;"androidx.camera"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;name&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;"camera-core"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;version&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;ref&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;"camerax"&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="na"&gt;androidx-camera-camera2&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;group&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;"androidx.camera"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;name&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;"camera-camera2"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;version&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;ref&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;"camerax"&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="na"&gt;androidx-camera-lifecycle&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;group&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;"androidx.camera"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;name&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;"camera-lifecycle"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;version&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;ref&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;"camerax"&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="na"&gt;androidx-camera-view&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;group&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;"androidx.camera"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;name&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;"camera-view"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;version&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;ref&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;"camerax"&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="na"&gt;google-ai-edge-litert&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;group&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;"com.google.ai.edge.litert"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;name&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;"litert"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;version&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;ref&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;"litert"&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;In &lt;code&gt;app/build.gradle.kts&lt;/code&gt;, add the dependencies &lt;strong&gt;and&lt;/strong&gt; disable compression for &lt;code&gt;.tflite&lt;/code&gt; assets so Android can memory-map (&lt;code&gt;mmap&lt;/code&gt;) the model file directly without decompressing it into RAM:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight kotlin"&gt;&lt;code&gt;&lt;span class="n"&gt;android&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="c1"&gt;// ...&lt;/span&gt;
    &lt;span class="n"&gt;androidResources&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="n"&gt;noCompress&lt;/span&gt; &lt;span class="o"&gt;+=&lt;/span&gt; &lt;span class="s2"&gt;"tflite"&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;

&lt;span class="n"&gt;dependencies&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="nf"&gt;implementation&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;libs&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;androidx&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;camera&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;core&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="nf"&gt;implementation&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;libs&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;androidx&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;camera&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;camera2&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="nf"&gt;implementation&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;libs&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;androidx&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;camera&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;lifecycle&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="nf"&gt;implementation&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;libs&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;androidx&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;camera&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;view&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="nf"&gt;implementation&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;libs&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;google&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;ai&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;edge&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;litert&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Tip:&lt;/strong&gt; Never skip &lt;code&gt;noCompress += "tflite"&lt;/code&gt;. If Gradle compresses your &lt;code&gt;.tflite&lt;/code&gt; asset inside the APK, &lt;code&gt;assets.openFd(...)&lt;/code&gt; will fail and prevent zero-copy memory mapping.&lt;/p&gt;
&lt;/blockquote&gt;




&lt;h2&gt;
  
  
  Step 2: Bundle Your Offline &lt;code&gt;.tflite&lt;/code&gt; Model and Labels
&lt;/h2&gt;

&lt;p&gt;Place your model and label files inside &lt;code&gt;app/src/main/assets/&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;app/src/main/assets/
├── vision_classifier_food.tflite   # aiy/vision/classifier/food_V1 (2,024 classes)
└── food_labels.txt                 # 2,024 class labels matching output indices
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Before writing Kotlin code, inspect your model's input and output tensor specs:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Input:&lt;/strong&gt; &lt;code&gt;[1, 192, 192, 3]&lt;/code&gt; (&lt;code&gt;UINT8&lt;/code&gt; RGB pixels &lt;code&gt;0..255&lt;/code&gt;)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Output:&lt;/strong&gt; &lt;code&gt;[1, 2024]&lt;/code&gt; (&lt;code&gt;UINT8&lt;/code&gt; quantized softmax probabilities)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Knowing whether your model uses &lt;code&gt;UINT8&lt;/code&gt; or &lt;code&gt;FLOAT32&lt;/code&gt; is critical — feeding &lt;code&gt;FLOAT32&lt;/code&gt; normalized data to a &lt;code&gt;UINT8&lt;/code&gt; model (or forgetting to dequantize &lt;code&gt;UINT8&lt;/code&gt; outputs) is one of the most common reasons an offline model outputs garbage.&lt;/p&gt;




&lt;h2&gt;
  
  
  Step 3: Initialize LiteRT (&lt;code&gt;CompiledModel&lt;/code&gt; + &lt;code&gt;Interpreter&lt;/code&gt; Fallback)
&lt;/h2&gt;

&lt;p&gt;LiteRT 2.1.0 provides the modern &lt;code&gt;CompiledModel&lt;/code&gt; + &lt;code&gt;TensorBuffer&lt;/code&gt; API (&lt;code&gt;com.google.ai.edge.litert.*&lt;/code&gt;) as well as a compatible &lt;code&gt;Interpreter&lt;/code&gt; (&lt;code&gt;org.tensorflow.lite.Interpreter&lt;/code&gt; bundled inside the &lt;code&gt;litert&lt;/code&gt; artifact).&lt;/p&gt;

&lt;p&gt;Older &lt;code&gt;.tflite&lt;/code&gt; models (like &lt;code&gt;aiy/vision/classifier/food_V1&lt;/code&gt;) sometimes lack &lt;code&gt;SignatureDefs&lt;/code&gt; or use &lt;code&gt;UINT8&lt;/code&gt; quantization parameters (&lt;code&gt;scale&lt;/code&gt; and &lt;code&gt;zeroPoint&lt;/code&gt;). Initializing &lt;code&gt;CompiledModel&lt;/code&gt; first and keeping &lt;code&gt;Interpreter&lt;/code&gt; ready for tensor metadata &amp;amp; fallback gives you the best of both worlds:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight kotlin"&gt;&lt;code&gt;&lt;span class="k"&gt;class&lt;/span&gt; &lt;span class="nc"&gt;LiteRtFoodClassifier&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
    &lt;span class="k"&gt;private&lt;/span&gt; &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;context&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nc"&gt;Context&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="k"&gt;private&lt;/span&gt; &lt;span class="kd"&gt;var&lt;/span&gt; &lt;span class="n"&gt;compiledModel&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nc"&gt;CompiledModel&lt;/span&gt;&lt;span class="o"&gt;?&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;null&lt;/span&gt;
    &lt;span class="k"&gt;private&lt;/span&gt; &lt;span class="kd"&gt;var&lt;/span&gt; &lt;span class="n"&gt;inputBuffers&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;List&lt;/span&gt;&lt;span class="o"&gt;&amp;lt;&lt;/span&gt;&lt;span class="nc"&gt;TensorBuffer&lt;/span&gt;&lt;span class="o"&gt;&amp;gt;?&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;null&lt;/span&gt;
    &lt;span class="k"&gt;private&lt;/span&gt; &lt;span class="kd"&gt;var&lt;/span&gt; &lt;span class="n"&gt;outputBuffers&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;List&lt;/span&gt;&lt;span class="o"&gt;&amp;lt;&lt;/span&gt;&lt;span class="nc"&gt;TensorBuffer&lt;/span&gt;&lt;span class="o"&gt;&amp;gt;?&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;null&lt;/span&gt;

    &lt;span class="k"&gt;private&lt;/span&gt; &lt;span class="kd"&gt;var&lt;/span&gt; &lt;span class="n"&gt;interpreter&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nc"&gt;Interpreter&lt;/span&gt;&lt;span class="o"&gt;?&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;null&lt;/span&gt;
    &lt;span class="k"&gt;private&lt;/span&gt; &lt;span class="kd"&gt;var&lt;/span&gt; &lt;span class="n"&gt;inputWidth&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;192&lt;/span&gt;
    &lt;span class="k"&gt;private&lt;/span&gt; &lt;span class="kd"&gt;var&lt;/span&gt; &lt;span class="n"&gt;inputHeight&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;192&lt;/span&gt;
    &lt;span class="k"&gt;private&lt;/span&gt; &lt;span class="kd"&gt;var&lt;/span&gt; &lt;span class="n"&gt;inputDataType&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;DataType&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="no"&gt;UINT8&lt;/span&gt;
    &lt;span class="k"&gt;private&lt;/span&gt; &lt;span class="kd"&gt;var&lt;/span&gt; &lt;span class="n"&gt;outputClassCount&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;2024&lt;/span&gt;
    &lt;span class="k"&gt;private&lt;/span&gt; &lt;span class="kd"&gt;var&lt;/span&gt; &lt;span class="n"&gt;outputDataType&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;DataType&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="no"&gt;UINT8&lt;/span&gt;
    &lt;span class="k"&gt;private&lt;/span&gt; &lt;span class="kd"&gt;var&lt;/span&gt; &lt;span class="n"&gt;outputScale&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;1f&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="mi"&gt;255f&lt;/span&gt;
    &lt;span class="k"&gt;private&lt;/span&gt; &lt;span class="kd"&gt;var&lt;/span&gt; &lt;span class="n"&gt;outputZeroPoint&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;
    &lt;span class="k"&gt;private&lt;/span&gt; &lt;span class="kd"&gt;var&lt;/span&gt; &lt;span class="n"&gt;labels&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;List&lt;/span&gt;&lt;span class="o"&gt;&amp;lt;&lt;/span&gt;&lt;span class="kt"&gt;String&lt;/span&gt;&lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;emptyList&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

    &lt;span class="nd"&gt;@Synchronized&lt;/span&gt;
    &lt;span class="k"&gt;fun&lt;/span&gt; &lt;span class="nf"&gt;ensureInitialized&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;Boolean&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;compiledModel&lt;/span&gt; &lt;span class="o"&gt;!=&lt;/span&gt; &lt;span class="k"&gt;null&lt;/span&gt; &lt;span class="o"&gt;||&lt;/span&gt; &lt;span class="n"&gt;interpreter&lt;/span&gt; &lt;span class="o"&gt;!=&lt;/span&gt; &lt;span class="k"&gt;null&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; &lt;span class="n"&gt;labels&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;isNotEmpty&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
            &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;
        &lt;span class="p"&gt;}&lt;/span&gt;

        &lt;span class="n"&gt;labels&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;loadLabelsFromAssets&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

        &lt;span class="c1"&gt;// 1. Modern LiteRT CompiledModel API&lt;/span&gt;
        &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;compiledReady&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;runCatching&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
            &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;model&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;CompiledModel&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;create&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
                &lt;span class="n"&gt;context&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;assets&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
                &lt;span class="s2"&gt;"vision_classifier_food.tflite"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
                &lt;span class="nc"&gt;CompiledModel&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;Options&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nc"&gt;Accelerator&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="no"&gt;CPU&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
            &lt;span class="p"&gt;)&lt;/span&gt;
            &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;inputs&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;createInputBuffers&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
            &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;outputs&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;createOutputBuffers&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
            &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;inputs&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;isNotEmpty&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; &lt;span class="n"&gt;outputs&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;isNotEmpty&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
                &lt;span class="n"&gt;compiledModel&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;model&lt;/span&gt;
                &lt;span class="n"&gt;inputBuffers&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;inputs&lt;/span&gt;
                &lt;span class="n"&gt;outputBuffers&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;outputs&lt;/span&gt;
                &lt;span class="kc"&gt;true&lt;/span&gt;
            &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="k"&gt;else&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
                &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;close&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
                &lt;span class="kc"&gt;false&lt;/span&gt;
            &lt;span class="p"&gt;}&lt;/span&gt;
        &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;getOrDefault&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

        &lt;span class="c1"&gt;// 2. Inspect tensor shapes, DataTypes, and quantization params via Interpreter&lt;/span&gt;
        &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;interpreterReady&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;runCatching&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
            &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;modelBuffer&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;loadMappedModelBuffer&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"vision_classifier_food.tflite"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
            &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;tflite&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;Interpreter&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;modelBuffer&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;Interpreter&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;Options&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;apply&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="nf"&gt;setNumThreads&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;4&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

            &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;inTensor&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;tflite&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;getInputTensor&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
            &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;inShape&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;inTensor&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;shape&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="c1"&gt;// [1, 192, 192, 3]&lt;/span&gt;
            &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;inShape&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;size&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;=&lt;/span&gt; &lt;span class="mi"&gt;3&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
                &lt;span class="n"&gt;inputHeight&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;inShape&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
                &lt;span class="n"&gt;inputWidth&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;inShape&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
            &lt;span class="p"&gt;}&lt;/span&gt;
            &lt;span class="n"&gt;inputDataType&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;inTensor&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;dataType&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

            &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;outTensor&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;tflite&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;getOutputTensor&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
            &lt;span class="n"&gt;outputClassCount&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;outTensor&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;shape&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;lastOrNull&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;?:&lt;/span&gt; &lt;span class="mi"&gt;2024&lt;/span&gt;
            &lt;span class="n"&gt;outputDataType&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;outTensor&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;dataType&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

            &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;qParams&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;outTensor&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;quantizationParams&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
            &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;qParams&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;scale&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="mi"&gt;0f&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
                &lt;span class="n"&gt;outputScale&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;qParams&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;scale&lt;/span&gt;
                &lt;span class="n"&gt;outputZeroPoint&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;qParams&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;zeroPoint&lt;/span&gt;
            &lt;span class="p"&gt;}&lt;/span&gt;

            &lt;span class="n"&gt;interpreter&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;tflite&lt;/span&gt;
            &lt;span class="kc"&gt;true&lt;/span&gt;
        &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;getOrDefault&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

        &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="n"&gt;compiledReady&lt;/span&gt; &lt;span class="o"&gt;||&lt;/span&gt; &lt;span class="n"&gt;interpreterReady&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Memory-mapping the model from &lt;code&gt;assets&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight kotlin"&gt;&lt;code&gt;&lt;span class="k"&gt;private&lt;/span&gt; &lt;span class="k"&gt;fun&lt;/span&gt; &lt;span class="nf"&gt;loadMappedModelBuffer&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;assetName&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;String&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nc"&gt;ByteBuffer&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="n"&gt;runCatching&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="n"&gt;context&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;assets&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;openFd&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;assetName&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;use&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;fd&lt;/span&gt; &lt;span class="o"&gt;-&amp;gt;&lt;/span&gt;
            &lt;span class="nc"&gt;FileInputStream&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;fd&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;fileDescriptor&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;channel&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;use&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;channel&lt;/span&gt; &lt;span class="o"&gt;-&amp;gt;&lt;/span&gt;
                &lt;span class="n"&gt;channel&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;map&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
                    &lt;span class="nc"&gt;FileChannel&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;MapMode&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="no"&gt;READ_ONLY&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
                    &lt;span class="n"&gt;fd&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;startOffset&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
                    &lt;span class="n"&gt;fd&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;declaredLength&lt;/span&gt;
                &lt;span class="p"&gt;)&lt;/span&gt;
            &lt;span class="p"&gt;}&lt;/span&gt;
        &lt;span class="p"&gt;}&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;getOrElse&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;bytes&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;context&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;assets&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;open&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;assetName&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;use&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;it&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;readBytes&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;
        &lt;span class="nc"&gt;ByteBuffer&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;allocateDirect&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;bytes&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;size&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;apply&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
            &lt;span class="nf"&gt;order&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nc"&gt;ByteOrder&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;nativeOrder&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
            &lt;span class="nf"&gt;put&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;bytes&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
            &lt;span class="nf"&gt;rewind&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
        &lt;span class="p"&gt;}&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;






&lt;h2&gt;
  
  
  Step 4: Preprocess Bitmaps &amp;amp; Run Inference
&lt;/h2&gt;

&lt;p&gt;Before passing a camera &lt;code&gt;Bitmap&lt;/code&gt; into LiteRT:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Convert &lt;code&gt;Bitmap.Config.HARDWARE&lt;/code&gt; to &lt;code&gt;ARGB_8888&lt;/code&gt; if needed (since &lt;code&gt;getPixels&lt;/code&gt; cannot read hardware bitmaps).&lt;/li&gt;
&lt;li&gt;Center-crop to a square and scale to &lt;code&gt;192×192&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Extract packed RGB bytes (&lt;code&gt;R&lt;/code&gt;, &lt;code&gt;G&lt;/code&gt;, &lt;code&gt;B&lt;/code&gt;) in row-major order.
&lt;/li&gt;
&lt;/ol&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight kotlin"&gt;&lt;code&gt;&lt;span class="k"&gt;private&lt;/span&gt; &lt;span class="k"&gt;fun&lt;/span&gt; &lt;span class="nf"&gt;centerCropAndScale&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;source&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nc"&gt;Bitmap&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;targetWidth&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;Int&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;targetHeight&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;Int&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nc"&gt;Bitmap&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;safeBitmap&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;source&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;config&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="nc"&gt;Bitmap&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;Config&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="no"&gt;HARDWARE&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="n"&gt;source&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;copy&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nc"&gt;Bitmap&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;Config&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="no"&gt;ARGB_8888&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="k"&gt;else&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="n"&gt;source&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;
    &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;minDim&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;min&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;safeBitmap&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;width&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;safeBitmap&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;height&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;cropX&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;safeBitmap&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;width&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="n"&gt;minDim&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="mi"&gt;2&lt;/span&gt;
    &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;cropY&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;safeBitmap&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;height&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="n"&gt;minDim&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="mi"&gt;2&lt;/span&gt;
    &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;cropped&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;Bitmap&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;createBitmap&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;safeBitmap&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;cropX&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;cropY&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;minDim&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;minDim&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;cropped&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;width&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="n"&gt;targetWidth&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; &lt;span class="n"&gt;cropped&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;height&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="n"&gt;targetHeight&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="n"&gt;cropped&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="k"&gt;else&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="nc"&gt;Bitmap&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;createScaledBitmap&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;cropped&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;targetWidth&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;targetHeight&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;When reading quantized &lt;code&gt;UINT8&lt;/code&gt; output bytes in Kotlin/Java, remember that &lt;code&gt;Byte&lt;/code&gt; is signed (&lt;code&gt;-128..127&lt;/code&gt;), so always mask with &lt;code&gt;and 0xFF&lt;/code&gt; and apply the quantization formula:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight kotlin"&gt;&lt;code&gt;&lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;probability&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;rawByte&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;toInt&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="mh"&gt;0xFF&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="n"&gt;outputZeroPoint&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="n"&gt;outputScale&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Running inference with &lt;code&gt;CompiledModel&lt;/code&gt; + &lt;code&gt;TensorBuffer&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight kotlin"&gt;&lt;code&gt;&lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;model&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;compiledModel&lt;/span&gt;
&lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;inputs&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;inputBuffers&lt;/span&gt;
&lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;outputs&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;outputBuffers&lt;/span&gt;
&lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;model&lt;/span&gt; &lt;span class="o"&gt;!=&lt;/span&gt; &lt;span class="k"&gt;null&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; &lt;span class="n"&gt;inputs&lt;/span&gt; &lt;span class="o"&gt;!=&lt;/span&gt; &lt;span class="k"&gt;null&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; &lt;span class="n"&gt;outputs&lt;/span&gt; &lt;span class="o"&gt;!=&lt;/span&gt; &lt;span class="k"&gt;null&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;probs&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;runCatching&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;rgbBytes&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;bitmapToRgbByteArray&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;croppedBitmap&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;inputWidth&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;inputHeight&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
        &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;inputDataType&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="nc"&gt;DataType&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="no"&gt;FLOAT32&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
            &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;floatIn&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;FloatArray&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;rgbBytes&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;size&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;i&lt;/span&gt; &lt;span class="o"&gt;-&amp;gt;&lt;/span&gt;
                &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;rgbBytes&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;i&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;toInt&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="mh"&gt;0xFF&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="mf"&gt;127.5f&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="mf"&gt;1.0f&lt;/span&gt;
            &lt;span class="p"&gt;}&lt;/span&gt;
            &lt;span class="n"&gt;inputs&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;writeFloat&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;floatIn&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
            &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;run&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;inputs&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;outputs&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
            &lt;span class="n"&gt;outputs&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;readFloat&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
        &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="k"&gt;else&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
            &lt;span class="n"&gt;inputs&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;writeInt8&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;rgbBytes&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
            &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;run&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;inputs&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;outputs&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
            &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;outputDataType&lt;/span&gt; &lt;span class="o"&gt;==&lt;/span&gt; &lt;span class="nc"&gt;DataType&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="no"&gt;FLOAT32&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
                &lt;span class="n"&gt;outputs&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;readFloat&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
            &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="k"&gt;else&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
                &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;rawOut&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;outputs&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;readInt8&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
                &lt;span class="nc"&gt;FloatArray&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;rawOut&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;size&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;i&lt;/span&gt; &lt;span class="o"&gt;-&amp;gt;&lt;/span&gt;
                    &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;rawOut&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;i&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;toInt&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="mh"&gt;0xFF&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="n"&gt;outputZeroPoint&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="n"&gt;outputScale&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;coerceAtLeast&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;0f&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
                &lt;span class="p"&gt;}&lt;/span&gt;
            &lt;span class="p"&gt;}&lt;/span&gt;
        &lt;span class="p"&gt;}&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;getOrNull&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Always release &lt;code&gt;TensorBuffer&lt;/code&gt;, &lt;code&gt;CompiledModel&lt;/code&gt;, and &lt;code&gt;Interpreter&lt;/code&gt; resources when your &lt;code&gt;ViewModel&lt;/code&gt; clears (&lt;code&gt;onCleared()&lt;/code&gt;).&lt;/p&gt;




&lt;h2&gt;
  
  
  Step 5: Stream Live Frames from CameraX
&lt;/h2&gt;

&lt;p&gt;In &lt;code&gt;CameraPreviewView.kt&lt;/code&gt;, configure &lt;code&gt;ImageAnalysis&lt;/code&gt; with &lt;code&gt;OUTPUT_IMAGE_FORMAT_RGBA_8888&lt;/code&gt; and &lt;code&gt;STRATEGY_KEEP_ONLY_LATEST&lt;/code&gt; so slow frames never queue up. Don't forget to rotate the &lt;code&gt;Bitmap&lt;/code&gt; using &lt;code&gt;imageProxy.imageInfo.rotationDegrees&lt;/code&gt; — otherwise your model receives sideways images and accuracy drops sharply:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight kotlin"&gt;&lt;code&gt;&lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;imageAnalysis&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;ImageAnalysis&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;Builder&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;setBackpressureStrategy&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nc"&gt;ImageAnalysis&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="no"&gt;STRATEGY_KEEP_ONLY_LATEST&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;setOutputImageFormat&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nc"&gt;ImageAnalysis&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="no"&gt;OUTPUT_IMAGE_FORMAT_RGBA_8888&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;build&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;

&lt;span class="n"&gt;imageAnalysis&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;setAnalyzer&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;analysisExecutor&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;imageProxy&lt;/span&gt; &lt;span class="o"&gt;-&amp;gt;&lt;/span&gt;
    &lt;span class="k"&gt;try&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;rawBitmap&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;imageProxy&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;toBitmap&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
        &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;rotation&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;imageProxy&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;imageInfo&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;rotationDegrees&lt;/span&gt;
        &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;uprightBitmap&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;rotation&lt;/span&gt; &lt;span class="o"&gt;!=&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
            &lt;span class="kd"&gt;val&lt;/span&gt; &lt;span class="n"&gt;matrix&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;Matrix&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;apply&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="nf"&gt;postRotate&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;rotation&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;toFloat&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;
            &lt;span class="nc"&gt;Bitmap&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;createBitmap&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
                &lt;span class="n"&gt;rawBitmap&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;rawBitmap&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;width&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;rawBitmap&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;height&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;matrix&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;
            &lt;span class="p"&gt;)&lt;/span&gt;
        &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="k"&gt;else&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
            &lt;span class="n"&gt;rawBitmap&lt;/span&gt;
        &lt;span class="p"&gt;}&lt;/span&gt;
        &lt;span class="nf"&gt;onFrameBitmap&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;uprightBitmap&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="k"&gt;finally&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="n"&gt;imageProxy&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;close&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;






&lt;h2&gt;
  
  
  Step 6: Stabilize Live Detections in the ViewModel
&lt;/h2&gt;

&lt;p&gt;Running inference on every camera frame can cause flickering or stale UI states if you aren't careful. Here are three practical rules we applied in &lt;code&gt;ScanFoodViewModel&lt;/code&gt;:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Throttle &amp;amp; guard concurrency:&lt;/strong&gt; Skip incoming frames if less than &lt;code&gt;850ms&lt;/code&gt; has passed or if an &lt;code&gt;AtomicBoolean&lt;/code&gt; (&lt;code&gt;isClassifyingFrame&lt;/code&gt;) is already active. Run inference on &lt;code&gt;Dispatchers.Default&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Use honest confidence thresholds:&lt;/strong&gt; A 2,024-class softmax spreads probability across many labels, so real top-1 confidence on a good match is often &lt;code&gt;28%–65%&lt;/code&gt;, not &lt;code&gt;90%+&lt;/code&gt;. Never artificially inflate confidence scores in code; instead calibrate your lock threshold (&lt;code&gt;MIN_LIVE_CONFIDENCE_PERCENT = 28&lt;/code&gt;) to the model's real distribution.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Smooth transitions instead of hard-locking:&lt;/strong&gt;

&lt;ul&gt;
&lt;li&gt;First match (&lt;code&gt;&amp;gt;= 28%&lt;/code&gt;) updates the UI right away.&lt;/li&gt;
&lt;li&gt;Switching to a &lt;strong&gt;different&lt;/strong&gt; food requires &lt;strong&gt;2 consecutive frames&lt;/strong&gt; (&lt;code&gt;pendingCandidateStreak &amp;gt;= 2&lt;/code&gt;) so brief camera shake doesn't overwrite a good detection.&lt;/li&gt;
&lt;li&gt;Pointing away from food (&lt;code&gt;&amp;lt; 18%&lt;/code&gt; for &lt;strong&gt;3 consecutive frames&lt;/strong&gt;) clears the previous item so the user can seamlessly scan a new plate.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;/ol&gt;




&lt;h2&gt;
  
  
  Step 7: Map Detected Labels to Offline Nutrition Data
&lt;/h2&gt;

&lt;p&gt;Once LiteRT returns a food label (e.g., &lt;code&gt;"Mango Sticky Rice"&lt;/code&gt;, &lt;code&gt;"Breakfast Burrito"&lt;/code&gt;), &lt;code&gt;FoodNutritionEstimator&lt;/code&gt; maps the label offline to calories, protein/carbs/fat, serving weight, and micronutrients (fiber, sugar, sodium, potassium, calcium, iron, vitamin C, cholesterol). Users can also tap any macro card, adjust serving multipliers (&lt;code&gt;0.5x&lt;/code&gt;, &lt;code&gt;1.0x&lt;/code&gt;, &lt;code&gt;1.5x&lt;/code&gt;, or custom grams), or switch between the top-3 predicted matches before logging to Room/SQLite.&lt;/p&gt;




&lt;h2&gt;
  
  
  Key Takeaways &amp;amp; Common Pitfalls
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Always inspect your &lt;code&gt;.tflite&lt;/code&gt; tensors:&lt;/strong&gt; Check input size (&lt;code&gt;192×192&lt;/code&gt;), channel order (&lt;code&gt;RGB&lt;/code&gt;), and quantization (&lt;code&gt;UINT8&lt;/code&gt; vs &lt;code&gt;FLOAT32&lt;/code&gt;) before writing preprocessing code.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Avoid overriding fine-grained classifiers with coarse segmenters:&lt;/strong&gt; If you combine a classifier with a semantic segmenter, let the whole-image classifier name cohesive dishes instead of forcing segment categories into hardcoded fallback names.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Reuse direct &lt;code&gt;ByteBuffer&lt;/code&gt;s for heavy outputs:&lt;/strong&gt; Large segmentation tensors (like &lt;code&gt;[1, 513, 513, 26] FLOAT32&lt;/code&gt; ≈ 27 MB) should reuse a cached direct &lt;code&gt;ByteBuffer&lt;/code&gt; rather than allocating per frame.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Keep live scanning responsive:&lt;/strong&gt; Use &lt;code&gt;STRATEGY_KEEP_ONLY_LATEST&lt;/code&gt;, throttle frames, and use short multi-frame streaks instead of permanently pausing the camera after the first detection.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;With LiteRT + CameraX, you get a fast, privacy-first, zero-cost AI scanner that works anywhere — even in airplane mode.&lt;br&gt;
Github: &lt;a href="https://github.com/tusharahmed11/NutriLens" rel="noopener noreferrer"&gt;https://github.com/tusharahmed11/NutriLens&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>litert</category>
      <category>machinelearning</category>
      <category>ondeviceai</category>
    </item>
  </channel>
</rss>
