NAFNet (Nonlinear Activation Free Network, ECCV 2022) image restoration, converted to LiteRT and running fully on the CompiledModel GPU (ML Drift) on Android. This is the SIDD-width32 variant — real-image denoising. NAFNet is a U-Net of NAFBlocks with no activation functions (SimpleGate = channel-split multiply), so the whole network is a clean CNN on the GPU.

A complete Android sample (image picker + before/after) is in the official google-ai-edge/litert-samples repo under compiled_model_api/image_restoration.

Pure CNN (no activations). Three numerically-exact re-authorings, the headline being SafeLayerNorm: NAFNet's residual stream grows large (|x|≈175 at the bottleneck), so the LayerNorm channel reductions Σ_c x and Σ_c (x−μ)² (~15M) overflow fp16 (max 65504) on the Mali delegate (which computes in fp16 regardless of the model dtype) → a grid artifact. Doing the reductions in a down-scaled x/S domain (S=128) and rescaling is exact and fp16-safe. Plus the Simplified Channel Attention AdaptiveAvgPool2d(1) → mean(3).mean(2), and the upsample Conv2d(1×1)+PixelShuffle(2) → depth-to-space ZeroStuffConvT2d.

Result: banned ops NONE, all tensors ≤4D, tflite-vs-torch corr 1.0, device-vs-torch corr 1.0.

A complete Android sample (image picker + before/after) is in the official google-ai-edge/litert-samples repo under compiled_model_api/image_restoration (push this .tflite in place of the deblur model).

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.

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.

XNNPACK declines these fp16 graphs — it reports failed to delegate DEPTHWISE_CONV_2D and then fails to allocate tensors — so there is no usable CPU number. Disabling XNNPACK falls back to reference kernels, which measured about 20× slower than the GPU on models of this size and would not represent CPU inference anyone would ship.

The NPU is 1.42x faster than the GPU (26.30 ms against 37.26 ms) and loads 7.14x faster (223 ms against 1594 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 0.72, 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. That first compile took 66 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.

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).

MIT. Upstream: megvii-research/NAFNet; weights NAFNet-SIDD-width32.