litert-community/U2Net-Portrait-Sketch-LiteRT

U²-Net Portrait — Photo → pencil line drawing (LiteRT GPU)

On-device portrait sketch generation running fully on the LiteRT CompiledModel GPU delegate (no CPU fallback). The U²-Net portrait model turns a face photo into a hand-drawn pencil line portrait — a fun creative / AR filter. ~12 ms/frame on a Pixel 8a.

• Architecture: U²-Net (RSU / nested residual U-blocks) — pure CNN.
• Weights: xuebinqin/U-2-Net (u2net_portrait) · Apache-2.0.
• Size: 176 MB.

Input (left) → generated pencil portrait (right). Photo: Unsplash (free license).

I/O

• Input: [1, 3, 512, 512] NCHW, RGB, x/max then ImageNet-normalized (mean [0.485,0.456,0.406], std [0.229,0.224,0.225]). A centered face works best.
• Output: [1, 1, 512, 512] in [0,1]. Min-max normalize, then invert (1 − x) for dark strokes on white paper.

GPU conversion

U²-Net is a pure CNN → fully GPU-compatible (893/893 nodes on the delegate, 1 partition; device corr 0.998683, ~12 ms) with one defensive patch: align_corners=True → False on the bilinear upsamples. CPU-exact vs PyTorch (corr 1.0).

Minimal usage

Kotlin (Android, LiteRT CompiledModel GPU)

val options = CompiledModel.Options(Accelerator.GPU)
val model = CompiledModel.create(context.assets, "portrait.tflite", options, null)
val inBufs = model.createInputBuffers()
val outBufs = model.createOutputBuffers()

inBufs[0].writeFloat(inputNCHW)          // [1,3,512,512] RGB, /max then ImageNet-norm
model.run(inBufs, outBufs)
val d = outBufs[0].readFloat()           // [512*512] 0..1; min-max normalize then 1-x -> pencil sketch

Python (LiteRT / ai-edge-litert)

import numpy as np
from ai_edge_litert.interpreter import Interpreter

it = Interpreter(model_path="portrait.tflite"); it.allocate_tensors()
inp, out = it.get_input_details(), it.get_output_details()
it.set_tensor(inp[0]["index"], x)        # [1,3,512,512] float32, RGB, /max, ImageNet-norm
it.invoke()
d = it.get_tensor(out[0]["index"])[0, 0]
d = (d - d.min()) / (d.max() - d.min()); sketch = 1.0 - d   # dark strokes on white

Conversion

Converted with litert-torch (build_portrait.py): loads the Apache-2.0 u2net_portrait weights and exports the sketch map.

License

Apache-2.0 (U²-Net / xuebinqin).