[AndyScherzinger]

MediaPipe Tasks sends usage statistics (task name, running mode, device info, latency and frame counts) to Google via datatransport/Clearcut on Android in every version, including 0.10.26; there is no option to turn it off. This replaces it with LiteRT running the selfie_segmenter.tflite model we already ship, so background effects work without that reporting.

• ImageSegmenterHelper (gplay) now runs the model with a LiteRT Interpreter on the CPU: the RGBA frame is scaled to the 256×256 model input and normalised, the person confidence is thresholded at 0.5 and scaled back to the frame size as the 0/1 mask BackgroundBlurGPUProcessor expects (1 = background, blurred). Frames arriving while an inference runs are dropped, like MediaPipe's live-stream mode. The public API is unchanged, so the blur pipeline is untouched.
• com.google.mediapipe:tasks-vision is removed together with its Renovate lock and verification metadata; com.google.ai.edge.litert:litert:1.4.2 is added (signed with the Google Maven key already trusted for androidx/Firebase). LiteRT 2.x is not used: it depends on Play Asset Delivery (com.google.android.play:ai-delivery), which F-Droid's scanner flags as non-free; Renovate is limited to <2.
• The generic and qa stubs only get their outdated comment fixed; enabling background effects in generic is a separate PR.

How the absence of telemetry was checked so far:

• LiteRT 1.4.2 litert and litert-api AARs: no datatransport, Firebase, Clearcut or Play references and no network code in the classes; the native library (all four ABIs) imports no socket, DNS, HTTP or TLS functions, needs only libc/libm/libdl/liblog and only refers to NNAPI and system libraries. No manifest components or permissions. Native libraries are 16 KB page aligned.
• Dependency tree (gplayReleaseRuntimeClasspath, master vs this branch): adds only litert and litert-api; removes tasks-vision, tasks-core, Flogger, protobuf-javalite and checker-compat-qual. genericReleaseRuntimeClasspath is unchanged.
• Merged gplay manifest: identical to master.
• gplay debug APK: no com.google.mediapipe or com.google.android.play.core classes; selfie_segmenter.tflite is stored uncompressed (needed for memory-mapping it).

🖼️ Screenshots

🏚️ Before	🏡 After
B	A

🚧 TODO

• Confirm mask polarity and quality on a device (derived from the shader, which blurs where the mask is non-zero)
• Parity check against the MediaPipe build: mask quality, inference time
• Network capture during a call with background blur, with a MediaPipe build as baseline that shows the Clearcut uploads
• Follow-up: the protobuf-javalite substitution in app/build.gradle.kts no longer matches anything in the release classpaths

🏁 Checklist

• ⛑️ Tests (unit and/or integration) are included or not needed
• 🔖 Capability is checked or not needed
• 🔙 Backport requests are created or not needed: /backport to stable-xx.x
• 📅 Milestone is set
• 🌸 PR title is meaningful (if it should be in the changelog: is it meaningful to users?)

🤖 AI (if applicable)

• The content of this PR was partly or fully generated using AI

[github-actions]

📱 QA build

Download	app-qa-debug.apk
QR code	Open the QR code for this download
Commit	275b001
Version	6984
Available until	7 days after this build

The QA build installs alongside a released Nextcloud app, so you can keep
using your existing install while testing.

Downloading the file requires a GitHub account, so open this link on the
device you want to test on, or transfer the APK to it.