A reproducible Unity project that measures how texture compression, audio compression, Korean font subsetting, and a runtime plugin change an Android APK.
The experiment builds eight APKs while changing one condition at a time, then repeats the full run. The maximum same-step difference between the two runs was 8 bytes.
Tested with Unity 6000.3.7f1, Android APK, IL2CPP, ARM64, Built-in Render Pipeline, Development Build off, and Managed Stripping Level Medium.
| Step | Controlled change | APK MiB | Delta MiB |
|---|---|---|---|
| 0 | Minimal experiment project | 14.32 | - |
| 1 | 2048x2048 texture, Android RGBA32 | 28.12 | +13.80 |
| 2 | Same texture, Android ASTC 6x6 | 16.14 | -11.98 |
| 3 | Add 30-second stereo audio, PCM | 21.19 | +5.05 |
| 4 | Same audio, Vorbis quality 0.5 | 16.60 | -4.59 |
| 5 | Add full Noto Sans KR dynamic font | 19.49 | +2.89 |
| 6 | Same font, app-specific subset | 16.63 | -2.86 |
| 7 | Add DOTween, preserved and referenced at runtime | 17.10 | +0.46 |
The exact numbers belong to this configuration. The reusable part is the method: fix the settings, change one variable, measure the final artifact, and keep the final APK size separate from Unity's BuildReport metrics.
- Unity
6000.3.7f1 - Android Build Support, SDK, NDK, and OpenJDK installed through Unity Hub
- macOS, Windows, or Linux supported by that Unity version
- Clone this repository.
- Open the repository root as a Unity project.
- Select
Hano Studio > Build Analyzer > Run Full Size Experiment. - Wait for all eight Android builds to finish.
Results are written to:
Builds/experiment-results.csvBuilds/Experiment/step-00-baseline.apkthroughstep-07-dotween-runtime.apkBuilds/Reports/*-summary.txtBuilds/Reports/*-assets.csv
git clone https://github.com/hellohanostudio/BuildAnalyzer.git
cd BuildAnalyzer
/Applications/Unity/Hub/Editor/6000.3.7f1/Unity.app/Contents/MacOS/Unity \
-batchmode \
-quit \
-projectPath "$(pwd)" \
-executeMethod HanoStudio.BuildAnalyzer.Editor.ExperimentBuildRunner.RunAllSteps \
-logFile "$(pwd)/Builds/full-experiment.log"Adjust the Unity executable path for Windows or Linux.
- One saved scene with one camera, one directional light, and one cube
- APK output, IL2CPP, ARM64 only
- Development Build off
- Managed Stripping Level Medium and engine code stripping on
- The same generated 2048x2048 high-entropy texture for RGBA32 and ASTC
- The same deterministic 30-second stereo WAV for PCM and Vorbis
- The same Noto Sans KR source font before and after subsetting
- The same DOTween Free DLL, excluded before step 7 and preserved at runtime in step 7
The scene contains a small runtime reference harness so each tested asset is genuinely referenced and survives normal build stripping.
The reporter records three values separately:
- Final APK file bytes from
FileInfo.Length BuildReport.summary.totalSize- The sum of
BuildReport.packedAssets[].contents[].packedSize
They are not interchangeable. In the final measured build they were 17.10 MiB, 143.54 MiB, and 5.60 MiB respectively.
The DOTween step shows why this matters. The APK grew by 0.46 MiB while the packed-asset report attributed only 192 bytes to the DLL path. Most growth appeared in IL2CPP native code and metadata.
Assets/BuildAnalyzer/Editor/
BuildAssetSizeReporter.cs # Exports ranked PackedAssetInfo data
ExperimentBuildRunner.cs # Generates inputs and builds all eight steps
ExperimentEvidenceWindow.cs # Displays verified results inside Unity
Assets/BuildAnalyzer/Runtime/
ExperimentReferences.cs # Keeps test inputs referenced at runtime
ExperimentInputs/
Fonts/ # Full and subset Noto Sans KR inputs
DOTween/ # Standard DOTween Free DLL and license
Results/
experiment-results-run1.csv
experiment-results-run2.csv
raw/ # Per-step summary and packed-asset exports
The original experiment code is licensed under the MIT License.
Noto Sans KR and DOTween keep their own licenses. See THIRD_PARTY_NOTICES.md and the license files under ExperimentInputs.