On-device background removal running fully on the LiteRT CompiledModel GPU
delegate (no CPU fallback). ormbg is a
fully open, Apache-2.0 foreground/alpha matte model (an ISNet trained for
photorealistic subject cut-out) — the permissively-licensed alternative to the
non-commercial RMBG-1.4. 246 ms per 1024² frame on a Pixel 8a (2026-09-05, run + readback).
Input:[1, 3, 1024, 1024] NCHW, RGB, x / 255 (no mean/std).
Output:[1, 1, 1024, 1024] alpha matte in [0,1]. Post-process: min-max
normalize, resize to the original size, use as the foreground alpha.
GPU conversion
ormbg is a pure CNN (ISNet RSU blocks). It converts fully GPU-compatible (246/246
nodes on the delegate, 1 partition; device corr 0.999881) with one
defensive patch: align_corners=True → False on the bilinear upsamples (the GPU
delegate rejects align_corners=True). CPU-exact vs PyTorch (corr 0.9999999999).
1implementation("com.google.ai.edge.litert:litert:2.2.0")// build.gradle.kts, plus androidResources { noCompress += "tflite" }23val remover = BgRemover.fromAssets(context)// once, off the main thread (~2.6 s: GPU compile + warm-up)4val cutout = remover.process(photo)?.cutout(photo)// Bitmap with a transparent background5remover.close()// on teardown; remover.cancel() aborts a call in flight
Verified 2026-09-05 on a Pixel 8a (Android 16, LiteRT 2.2.0): 246/246 nodes on the GPU, 246 ms per
1024² image (median of 20, run + readback, thermal status 0), 361 ms including pre/post-processing;
device output vs host CPU fp32 on the same input: corr 1.000000, max |diff| 0.00905. The recipe ships an
instrumented test that reproduces these values on your device. Provenance: converted with litert-torch
0.9.4 by john-rocky (build_ormbg.py,
re-run 2026-09-05, byte-identical to this file); recipe and measurements in
john-rocky/LiteRT-Models.
Conversion
Converted with litert-torch (build_ormbg.py): loads the ISNet weights and exports.
Performance
Measured on a Pixel 8a (Tensor G3, Android 16) with the standard TFLite benchmark_model tool — 10 warm-up runs then 50 timed runs, reported as the tool's mean.
Runtime
Backend
Graph on GPU
Latency
LiteRT CompiledModel (LITERT_CL)
GPU
246 / 246
246 ms (median of 20, run + readback, 2026-09-05)
TFLite benchmark_model (TfLiteGpuDelegateV2)
GPU (OpenCL)
246 / 246
244.4 ms
TFLite benchmark_model
CPU (XNNPACK, 4 threads)
—
5263.2 ms
The two GPU rows now agree: the classic delegate's 244.4 ms and CompiledModel's 246 ms are the same work. An earlier version of this card quoted ~10 ms for the CompiledModel row; that number timed the asynchronous run() call alone.
Snapdragon NPU (Hexagon)
This file runs on the Qualcomm Hexagon NPU as published — no conversion and no
pre-compiled artifact. LiteRT compiles it on the device and caches the result.
Measured on a physical Samsung Galaxy S26 (Snapdragon 8 Elite Gen 5 / SM8850,
Hexagon v81) with LiteRT CompiledModel 2.2.0 — 5 warm-up runs then 50 timed runs, one
accelerator per process, every row taken at device thermal status NONE.
Compute unit
Inference (median / min)
Load
Start headroom
NPU (Hexagon) — first launch
24.45 ms / 23.76 ms
11606 ms
0.60
NPU (Hexagon) — cached
23.95 ms / 23.85 ms
363 ms
0.62
GPU (Adreno)
74.42 ms / 73.44 ms
1178 ms
0.62
The NPU is 3.1x faster on inference here (23.95 ms against 74.42 ms). The first launch pays once for on-device compilation; every launch after that
loads in 363 ms against 1178 ms for the GPU (3.2x), because the GPU rebuilds its
shaders each time. The file is fp16 and needs no int8 quantization to reach the NPU.
Running it on the NPU
Put these in jniLibs/arm64-v8a/. None of them are distributed from this repository —
the first two come from Google, the rest from Qualcomm's own SDK:
Qualcomm QAIRT — the same zip ships fetch_qualcomm_library.sh, which downloads the SDK and copies them for you
Pick the runtime matching the device's Hexagon version: SM8550 → v73, SM8650 → v75,
SM8750 → v79, SM8850 → v81.
kotlin
1val env = Environment.create(2 context,3mapOf(4 Environment.Option.DispatchLibraryDir to context.applicationInfo.nativeLibraryDir,5// Required for on-device compilation. Without it the model silently runs on CPU.6 Environment.Option.CompilerPluginLibraryDir to context.applicationInfo.nativeLibraryDir,7),8)9val options = CompiledModel.Options(Accelerator.NPU).apply{10 qualcommOptions = CompiledModel.QualcommOptions(11 htpPerformanceMode = CompiledModel.QualcommOptions.HtpPerformanceMode.BURST
12)13}14val model = CompiledModel.create(context.assets,"ormbg.tflite", options, env)
Build settings: useLegacyPackaging = true under packaging { jniLibs { … } }, so the
DSP can open the skel from a real path, and Kotlin 2.3+ for LiteRT 2.2.0's metadata.
Every NPU failure here is silent. There is no error when the NPU is unavailable —
you get a plausible CPU number instead. Confirm from logcat which delegate took the
graph: Replacing 1 out of 1 node(s) with delegate (DispatchDelegate) is the NPU,
while ... (TfLiteXNNPackDelegate) is the CPU. A missing library is reported only as
a W-level dlopen failed line under a generic No compiler plugin found summary.
On the conditions. Thermal headroom is reported as measured, where 1.0 is the
throttling threshold. All rows were taken at a comparable headroom and compare directly;
figures taken at a different headroom will differ. Each accelerator ran in its own
process, because LiteRT's Environment is shared within one and the first model load
fixes the options for every later one.
Raspberry Pi 5 (CPU)
Measured on a Raspberry Pi 5 Model B Rev 1.1 (8 GB, Raspberry Pi OS 64-bit) with the LiteRT benchmark_model tool from litert-cli-nightly 0.2.0.dev20260805: CPU inference (XNNPACK, 4 threads), 3 invocations per file of 10 warm-up plus 50 timed runs (the tool caps a phase at 150 s, so very slow graphs run fewer — the Runs column is the actual timed total). The latency is the median across invocations; the spread is the min–max over all timed runs. No thermal throttling occurred during these runs (vcgencmd get_throttled stayed 0x0).
