On-device clothing segmentation running fully on the LiteRT CompiledModel GPU
delegate (no CPU fallback). cloth-segmentation
is a U²-Net trained on iMaterialist-Fashion to segment upper-body / lower-body /
full-body clothing — the building block for virtual try-on and fashion apps. ~88 ms/frame
on a Pixel 8a.
Architecture: U²-Net (RSU nested residual blocks), 4-class head — pure CNN.
Output:[1, 4, 768, 768] logits — argmax over the 4 classes: 0 = background,
1 = upper body, 2 = lower body, 3 = full body (dress).
GPU conversion
U²-Net is a pure CNN → fully GPU-compatible (254/254 nodes on the delegate, 1
partition; device corr 0.999798, ~88 ms) with one defensive patch: align_corners=True
→ False on the bilinear upsamples (the GPU delegate rejects align_corners=True).
CPU-exact vs PyTorch (corr 1.0).
Minimal usage
Kotlin (Android, LiteRT CompiledModel GPU)
kotlin
1val options = CompiledModel.Options(Accelerator.GPU)2val model = CompiledModel.create(context.assets,"clothseg.tflite", options,null)3val inBufs = model.createInputBuffers()4val outBufs = model.createOutputBuffers()56inBufs[0].writeFloat(inputNCHW)// [1,3,768,768] RGB, (x/255-0.5)/0.57model.run(inBufs, outBufs)8valout= outBufs[0].readFloat()// [4*768*768]; per pixel p argmax over the 4 class planes9// class 0 bg, 1 upper, 2 lower, 3 full-body
Converted with litert-torch (build_clothseg.py): loads the MIT U²-Net cloth weights
and exports the 4-class graph.
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
254 / 254
~88 ms
TFLite benchmark_model (TfLiteGpuDelegateV2)
GPU (OpenCL)
254 / 254
496.4 ms
The two GPU rows are different runtimes, not a contradiction. The LITERT_CL figure is the one recorded when this model shipped, taken through LiteRT's own CompiledModel accelerator — the path the Kotlin sample app and the LiteRT API use. The TfLiteGpuDelegateV2 figure is the classic TFLite OpenCL delegate, measured with a tool anyone can download and re-run. They agree on how much of the graph the GPU takes; they disagree on speed, and the classic delegate is the slower of the two here. Read the TfLiteGpuDelegateV2 row as a reproducible floor, not as this model's speed on LiteRT.
Snapdragon NPU (Hexagon)
The NPU is 2.29x faster than the GPU (65.47 ms against 149.9 ms) and loads 7.76x faster (247 ms against 1919 ms).
backend
inference (median / min)
load
NPU (Hexagon v81)
65.47 ms / 62.06 ms
247 ms
GPU (Adreno)
149.9 ms / 146.3 ms
1919 ms
Measured on a Samsung Galaxy S26 (Snapdragon 8 Elite Gen 5 / SM8850, Hexagon v81, Android 16), LiteRT CompiledModel 2.2.0, one accelerator per process, 5 warm-up runs then N=50 timed runs, median reported. Every run held thermal status NONE throughout. Headroom 0.54-0.61, where 1.0 is the throttling threshold.
The NPU rows here ran artifacts compiled ahead of time for SM8850 with QAIRT 2.47.0; the GPU rows ran the published files as they are. LiteRT can also compile for the NPU on the device at first load, which is what lets you ship the published file unchanged — that path and the ten runtime libraries it needs are in the NPU recipe, and we did not measure it here. GPU wiring is in the GPU recipe.
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).
File
Inference (median)
Spread (min–max)
Runs
Peak memory
clothseg.tflite
6,544.6 ms
6,414.9–6,715.9 ms
69
1823 MB
License
MIT (cloth-segmentation / levindabhi). Trained on iMaterialist-Fashion-2019.
