litert-community/YuNet-Face-LiteRT

YuNet — LiteRT (on-device face detection, fully-GPU)

YuNet (ShiqiYu/libfacedetection), a tiny fast face detector (faces + 5 landmarks), converted to LiteRT and running fully on the CompiledModel GPU (ML Drift) on Android. 0.076 M params / 0.3 MB fp16.

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

nodes on GPU	146 / 146 LITERT_CL (full residency)
inference	~4 ms (640×640)
size	0.3 MB (fp16)
accuracy	device-vs-PyTorch corr 0.9999 (all 12 outputs)

image[1,3,640,640] (BGR, 0-255) →[GPU: YuNet]→ 12 outputs: cls/obj/bbox/kps × strides {8,16,32}

How it converts (litert-torch) — clean, no re-authoring

Pure CNN (depthwise-separable ConvDPUnit) + a nearest-upsample neck (F.interpolate(mode="nearest") → RESIZE_NEAREST_NEIGHBOR, no transposed conv) + non-padded MaxPool2d (no PADV2). The head's per-stride permute/reshape/sigmoid is baked in → 12 decode-ready outputs. Banned ops NONE, ≤4D, tflite-vs-torch corr 1.0, device-vs-torch corr 0.9999.

Decode (host-side) & preprocessing

Preprocessing: letterbox to 640×640, BGR, 0-255, no normalization. Anchor-free priors (px=col·s, py=row·s, offset 0): score=cls·obj, box=center+exp(wh)·s, 5 landmarks kps·s+prior, then NMS.

Minimal usage

Android (Kotlin, CompiledModel GPU)

val model = CompiledModel.create(context.assets, "yunet_fp16.tflite",
    CompiledModel.Options(Accelerator.GPU), null)
val inputs = model.createInputBuffers(); val outputs = model.createOutputBuffers()
inputs[0].writeFloat(bgr)           // [1,3,640,640] NCHW BGR, 0-255 (no normalization)
model.run(inputs, outputs)
// 12 outputs in order: cls x3 [1,N,1], obj x3 [1,N,1], bbox x3 [1,N,4], kps x3 [1,N,10]
// for strides {8,16,32}, N = (640/stride)^2 = 6400/1600/400. Decode = Python below.
val cls8 = outputs[0].readFloat()

Python (desktop verification)

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

im = Image.open("faces.jpg").convert("RGB").resize((640, 640))
bgr = np.asarray(im, np.float32)[:, :, ::-1]                      # BGR, 0-255
x = bgr.transpose(2, 0, 1)[None].copy()                           # [1,3,640,640]

it = Interpreter(model_path="yunet_fp16.tflite"); it.allocate_tensors()
it.set_tensor(it.get_input_details()[0]["index"], x); it.invoke()
o = [it.get_tensor(d["index"])[0] for d in it.get_output_details()]
# output order: cls x3, obj x3, bbox x3, kps x3 (strides 8, 16, 32)
dets = []
for li, s in enumerate([8, 16, 32]):
    cls, obj, bb, kp = o[li][:, 0], o[3 + li][:, 0], o[6 + li], o[9 + li]
    fw = 640 // s
    for i in np.where(cls * obj > 0.6)[0]:                        # score threshold
        px, py = (i % fw) * s, (i // fw) * s
        cx, cy = bb[i, 0] * s + px, bb[i, 1] * s + py
        w, h = math.exp(bb[i, 2]) * s, math.exp(bb[i, 3]) * s
        lm = [(kp[i, 2 * j] * s + px, kp[i, 2 * j + 1] * s + py) for j in range(5)]
        dets.append(([cx - w/2, cy - h/2, cx + w/2, cy + h/2], float(cls[i] * obj[i]), lm))

def iou(a, b):                                                    # greedy NMS, IoU 0.45
    ix = max(0, min(a[2], b[2]) - max(a[0], b[0])); iy = max(0, min(a[3], b[3]) - max(a[1], b[1]))
    u = (a[2]-a[0])*(a[3]-a[1]) + (b[2]-b[0])*(b[3]-b[1]) - ix*iy
    return ix * iy / u if u > 0 else 0
dets.sort(key=lambda d: -d[1]); faces = []
for d in dets:
    if all(iou(d[0], f[0]) < 0.45 for f in faces): faces.append(d)
for box, score, lm in faces: print(f"face {score:.2f}", np.round(box, 1), "landmarks", np.round(lm, 1))

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	146 / 146	~4 ms
TFLite benchmark_model (TfLiteGpuDelegateV2)	GPU (OpenCL)	146 / 146	21.1 ms
TFLite benchmark_model	CPU (XNNPACK, 4 threads)	—	XNNPACK declined the graph

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.

Snapdragon NPU (Hexagon)

The NPU is 1.79x faster than the GPU (1.85 ms against 3.31 ms) and loads 5.75x faster (100 ms against 577 ms).

backend	compiled	inference (median / min)	load
NPU (Hexagon v81)	on-device JIT	1.85 ms / 1.80 ms	100 ms
GPU (Adreno)	—	3.31 ms / 2.20 ms	577 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 0.53, 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 979 ms 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
yunet_fp16.tflite	33.7 ms	33.4–35.2 ms	150	138 MB

License

BSD-3-Clause. Upstream: ShiqiYu/libfacedetection.