NIMA-LiteRT

NIMA — LiteRT on-device image quality assessment

NIMA (Neural Image Assessment) (idealo, Apache-2.0) re-authored for LiteRT: score a photo's quality on a 1-10 scale. Two MobileNet models — aesthetic (AVA) and technical (TID2013) — each predict a 10-bin score distribution; the score is the distribution mean. Both run fully on the CompiledModel GPU (~6.4 MB each).

Verified on a Pixel 8a: ~173 ms for both models; tflite-vs-Keras score parity 0.999998 (aesthetic) / 0.999915 (technical).

Files

image →[resize 224² · MobileNet /127.5−1]→ [GPU MobileNet]→ softmax dist[10] →[Σ i·pᵢ]→ score 1-10

Minimal usage (Python)

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

img = Image.open("photo.jpg").convert("RGB").resize((224, 224))
x = (np.asarray(img, np.float32) / 127.5 - 1.0)[None]        # NHWC, [-1,1]

def score(model):
    it = Interpreter(model_path=model); it.allocate_tensors()
    it.set_tensor(it.get_input_details()[0]["index"], x); it.invoke()
    dist = it.get_tensor(it.get_output_details()[0]["index"])[0]
    return float((np.arange(10) + 1) @ dist)                 # mean over 1..10

print("aesthetic", score("nima_aesthetic_fp16.tflite"))
print("technical", score("nima_technical_fp16.tflite"))

Minimal usage (Kotlin, LiteRT CompiledModel)

val m = CompiledModel.create(assets, "nima_aesthetic_fp16.tflite", CompiledModel.Options(Accelerator.GPU), null)
val inp = m.createInputBuffers(); val out = m.createOutputBuffers()
inp[0].writeFloat(preprocess(bitmap))       // resize 224², NHWC, v/127.5f - 1f
m.run(inp, out)
val dist = out[0].readFloat()               // [10]
var score = 0f; for (i in 0 until 10) score += (i + 1) * dist[i]   // 1-10

Upstream

idealo/image-quality-assessment (Apache-2.0) — NIMA MobileNet aesthetic + technical weights. Paper: NIMA: Neural Image Assessment (Talebi & Milanfar, 2018).

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.

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)

• nima_aesthetic_fp16.tflite — the NPU is 1.52x faster than the GPU (0.565 ms against 0.858 ms) and loads 3.61x faster (101 ms against 366 ms).

• nima_technical_fp16.tflite — the NPU is 1.54x faster than the GPU (0.554 ms against 0.854 ms) and loads 3.74x faster (99 ms against 371 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 0.67-0.67, 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).