litert-community/Fast-Neural-Style-LiteRT

Measured on device (edge-compat, style_candy_fp16): Galaxy S26 · LiteRT 2.2.0 · GPU (ML Drift) · 13.9 ms p50 (2026-08-25); Galaxy S26 · LiteRT 2.2.0 · NPU (QNN/HTP) · 7.41 ms p50 (2026-08-25); Raspberry Pi 5 · LiteRT 2.2.0.dev20260804 · CPU/XNNPACK, 4 threads · 287 ms p50 (2026-08-31); browser · Chromium 151 on M4 Max · LiteRT.js 2.5.3 · WebGPU · 7.34 ms p50 · output differs from CPU (max rel diff 0.05) (2026-08-11). Record: https://github.com/john-rocky/edge-compat/blob/main/cards/fast-neural-style__style_candy_fp16/CARD.md

Measured on device (edge-compat, style_mosaic_fp16): Galaxy S26 · LiteRT 2.2.0 · GPU (ML Drift) · 13.8 ms p50 (2026-08-25); Galaxy S26 · LiteRT 2.2.0 · NPU (QNN/HTP) · 7.41 ms p50 (2026-08-25); Raspberry Pi 5 · LiteRT 2.2.0.dev20260804 · CPU/XNNPACK, 4 threads · 288 ms p50 (2026-08-31); browser · Chromium 151 on M4 Max · LiteRT.js 2.5.3 · WebGPU · 7.19 ms p50 · output differs from CPU (max rel diff 0.016) (2026-08-11). Record: https://github.com/john-rocky/edge-compat/blob/main/cards/fast-neural-style__style_mosaic_fp16/CARD.md

Measured on device (edge-compat, style_rain_princess_fp16): Galaxy S26 · LiteRT 2.2.0 · GPU (ML Drift) · 13.8 ms p50 (2026-08-25); Galaxy S26 · LiteRT 2.2.0 · NPU (QNN/HTP) · 7.48 ms p50 (2026-08-25); Raspberry Pi 5 · LiteRT 2.2.0.dev20260804 · CPU/XNNPACK, 4 threads · 288 ms p50 (2026-08-31); browser · Chromium 151 on M4 Max · LiteRT.js 2.5.3 · WebGPU · 7.34 ms p50 · output differs from CPU (max rel diff 0.25) (2026-08-11). Record: https://github.com/john-rocky/edge-compat/blob/main/cards/fast-neural-style__style_rain_princess_fp16/CARD.md

Measured on device (edge-compat, style_udnie_fp16): Galaxy S26 · LiteRT 2.2.0 · GPU (ML Drift) · 13.8 ms p50 (2026-08-25); Galaxy S26 · LiteRT 2.2.0 · NPU (QNN/HTP) · 7.42 ms p50 (2026-08-25); Raspberry Pi 5 · LiteRT 2.2.0.dev20260804 · CPU/XNNPACK, 4 threads · 288 ms p50 (2026-08-31); browser · Chromium 151 on M4 Max · LiteRT.js 2.5.3 · WebGPU · 7.37 ms p50 · output differs from CPU (max rel diff 0.0041) (2026-08-11). Record: https://github.com/john-rocky/edge-compat/blob/main/cards/fast-neural-style__style_udnie_fp16/CARD.md

Fast Neural Style Transfer — LiteRT (on-device, fully-GPU, 4 styles)

Fast neural style transfer (PyTorch examples TransformerNet, Johnson et al.), converted to LiteRT and running fully on the CompiledModel GPU (ML Drift) on Android. Applies an artistic style to a photo — 4 styles (candy / mosaic / rain_princess / udnie), each a 3.5 MB fp16 graph.

On-device (Pixel 8a, Tensor G3 — verified)

nodes on GPU	350 / 350 LITERT_CL (full residency)
inference	~9 ms (256×256)
size	3.5 MB per style (fp16)
accuracy	device-vs-PyTorch corr 0.9998–0.9999 (all 4 styles)

image[1,3,256,256] (RGB 0-255) →[GPU: TransformerNet]→ stylized[1,3,256,256] (RGB 0-255)

Minimal usage

Android (Kotlin, CompiledModel GPU)

val model = CompiledModel.create(context.assets, "style_candy_fp16.tflite",
    CompiledModel.Options(Accelerator.GPU), null)
val inputs = model.createInputBuffers()
val outputs = model.createOutputBuffers()
inputs[0].writeFloat(chw)            // [1,3,256,256] RGB 0-255, NCHW
model.run(inputs, outputs)
val stylized = outputs[0].readFloat()    // [1,3,256,256] RGB 0-255

Python (desktop verification)

import numpy as np
from PIL import Image
from ai_edge_litert.interpreter import Interpreter

img = Image.open("photo.jpg").convert("RGB")
w, h = img.size; s = min(w, h)
img = img.crop(((w-s)//2, (h-s)//2, (w+s)//2, (h+s)//2)).resize((256, 256))
x = np.asarray(img, np.float32).transpose(2, 0, 1)[None]     # 0-255, no normalization

# candy / mosaic / rain_princess / udnie
it = Interpreter(model_path="style_candy_fp16.tflite"); it.allocate_tensors()
it.set_tensor(it.get_input_details()[0]["index"], x); it.invoke()
y = it.get_tensor(it.get_output_details()[0]["index"])[0]    # [3,256,256] RGB 0-255
Image.fromarray(y.transpose(1, 2, 0).clip(0, 255).astype(np.uint8)).save("stylized.png")

How it converts (litert-torch) — three numerically-exact re-authorings

1. ReflectionPad2d → zero-pad (GATHER_ND → PAD; border-only difference).
2. Large conv activations → conv-weight scaling. The conv outputs reach ≈ |5000|, where the Mali delegate's fp16 conv accumulation loses precision → garbage (device corr 0.34 at full residency — residency ≠ correctness). Each conv is followed by an InstanceNorm (which is scale-invariant), so scaling those conv weights down so the output is ≈ |10| is exact (IN output unchanged) and keeps the fp16 accumulation precise → corr 1.0.
3. InstanceNorm → SafeInstanceNorm (down-scaled-domain spatial reduction, fp16-safe; SafeLayerNorm class).

Upsample is interpolate(nearest) (no transposed conv → no ZeroStuff). Result: banned ops NONE, ≤4D, tflite-vs-torch corr 1.0, device-vs-torch corr 0.9999.

Preprocessing

Center-crop to square, resize to 256×256, RGB 0–255 (no normalization), NCHW. Output is 0–255 RGB (clamp).

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
TFLite benchmark_model (TfLiteGpuDelegateV2) — style_udnie_fp16.tflite	GPU (OpenCL)	350 / 350	37.4 ms
TFLite benchmark_model (TfLiteGpuDelegateV2) — style_mosaic_fp16.tflite	GPU (OpenCL)	350 / 350	37.5 ms
TFLite benchmark_model (TfLiteGpuDelegateV2) — style_rain_princess_fp16.tflite	GPU (OpenCL)	350 / 350	37.7 ms
TFLite benchmark_model (TfLiteGpuDelegateV2) — style_candy_fp16.tflite	GPU (OpenCL)	350 / 350	37.7 ms
TFLite benchmark_model — style_udnie_fp16.tflite	CPU (XNNPACK, 4 threads)	—	406.0 ms
TFLite benchmark_model — style_mosaic_fp16.tflite	CPU (XNNPACK, 4 threads)	—	404.3 ms
TFLite benchmark_model — style_rain_princess_fp16.tflite	CPU (XNNPACK, 4 threads)	—	404.5 ms
TFLite benchmark_model — style_candy_fp16.tflite	CPU (XNNPACK, 4 threads)	—	402.5 ms

Any on-device figure recorded when this model shipped came from a different runtime. It was taken through LiteRT's own CompiledModel accelerator (logcat reports it as LITERT_CL), which is the path the Kotlin sample app and the LiteRT API use, and it appears elsewhere on this card. The rows above are the classic TFLite OpenCL delegate, measured with a tool anyone can download and re-run. The two are not comparable, so read the rows above as a reproducible floor rather than as this model's speed on LiteRT.

Snapdragon NPU (Hexagon)

• style_candy_fp16.tflite — the NPU is 1.87x faster than the GPU (7.41 ms against 13.88 ms) and loads 5.85x faster (104 ms against 609 ms).
• style_mosaic_fp16.tflite — the NPU is 1.87x faster than the GPU (7.41 ms against 13.85 ms) and loads 5.78x faster (106 ms against 610 ms).
• style_rain_princess_fp16.tflite — the NPU is 1.85x faster than the GPU (7.48 ms against 13.85 ms) and loads 5.76x faster (105 ms against 603 ms).
• style_udnie_fp16.tflite — the NPU is 1.86x faster than the GPU (7.42 ms against 13.76 ms) and loads 5.78x faster (107 ms against 618 ms).

file	backend	compiled	inference (median / min)	load
style_candy_fp16.tflite	NPU (Hexagon v81)	on-device JIT	7.41 ms / 7.35 ms	104 ms
style_candy_fp16.tflite	GPU (Adreno)	—	13.88 ms / 13.33 ms	609 ms
style_mosaic_fp16.tflite	NPU (Hexagon v81)	on-device JIT	7.41 ms / 7.35 ms	106 ms
style_mosaic_fp16.tflite	GPU (Adreno)	—	13.85 ms / 13.52 ms	610 ms
style_rain_princess_fp16.tflite	NPU (Hexagon v81)	on-device JIT	7.48 ms / 7.35 ms	105 ms
style_rain_princess_fp16.tflite	GPU (Adreno)	—	13.85 ms / 13.48 ms	603 ms
style_udnie_fp16.tflite	NPU (Hexagon v81)	on-device JIT	7.42 ms / 7.33 ms	107 ms
style_udnie_fp16.tflite	GPU (Adreno)	—	13.76 ms / 13.36 ms	618 ms

Measured on a Samsung Galaxy S26 (Snapdragon 8 Elite Gen 5 / SM8850, Hexagon v81, Android 16) with 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 0.68–0.70, where 1.0 is the throttling threshold.

The NPU rows ran the published file unchanged. LiteRT compiled it for the Hexagon on the device at first load. Those first compiles took 2.8 s to 2.9 s here. The load column above is the cached load every later run pays. Recipe and the runtime libraries it needs: NPU guide.

GPU wiring: GPU guide.

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
style_candy_fp16.tflite	287.5 ms	285.8–289.9 ms	150	190 MB
style_mosaic_fp16.tflite	288.0 ms	285.8–290.6 ms	150	190 MB
style_rain_princess_fp16.tflite	287.6 ms	286.2–324.6 ms	150	190 MB
style_udnie_fp16.tflite	287.5 ms	286.1–292.1 ms	150	190 MB

License

BSD-3-Clause. Upstream: pytorch/examples.