upscale-factor-nano

46,899 parameters. 187 KB. Was this image upscaled, and by how much?

Given a 64×64 patch, predicts whether it was resampled up from a smaller original by 2×, 3×, or 4× — the signature of a "4K" stream that is really 720p wearing a bigger container.

Domain measured / deployment domain tested: measured on COCO val2017 photographs with real bicubic/bilinear resampling; deployment domain: no GAN or learned upscaler, and no camera ISP upscaling, tested. (Fifth line of the card standard, added 2026-09-02: a number is only as good as the domain it was measured in.)

Measured

Trained on real photographs (COCO val2017) with real resampling. Split by source image.

task	nano (held out)	best scalar (in-sample)	chance
upscale factor — 2× / 3× / 4×	0.980	0.489	0.344
upscaled at all, yes/no	0.977	0.847	0.514

The factor task is the one worth having. Energy loss alone cannot answer "by how much" — every tested scalar (mean, std, Laplacian variance, HF ratio, gradient, entropy) lands between 0.357 and 0.489, because the answer lives in the periodicity interpolation leaves behind, not in how much detail is missing. The model beats the best optimistically-fitted scalar by +0.491.

A prediction made before running this was wrong and is recorded here for honesty: the binary task was expected to fall to a scalar, since upscaling removes high frequencies much as blur does. It did not — 0.977 vs 0.847. Upscaling leaves structure a single statistic cannot capture even for the coarse question.

Validated on a second sensor (2026-08-25)

Everything above is COCO — one corpus, one kind of imagery. Tested on 1,002 patches from a Logitech BRIO watching a real office: a different sensor, a different image pipeline with its own denoising and sharpening, and frames that passed through the camera's own MJPG compression on the way out.

	accuracy	2x	3x	4x
COCO photographs (held out)	0.982	0.981	0.964	1.000
Logitech BRIO, real office	0.876	0.757	0.877	0.997

It transfers, losing about 10 points. The loss is entirely concentrated in the 2x class (0.981 → 0.757) while 4x is untouched (1.000 → 0.997), which is the physically expected shape: 2x upscaling leaves the subtlest periodic signature and lossy compression attenuates exactly those high frequencies first. This card previously listed "heavy compression after upscaling" as an untested risk; it is now measured, and the number to carry is 2x detection degrades first.

Flat patches (std < 6) were skipped, since sky and blank wall contain nothing to resample.

Limitation of this test: the BRIO offers a maximum of 1920x1080 over V4L2, so whether those frames are sensor-native or internally rescaled from a larger array could not be determined here. If they are rescaled, the real-world number is if anything pessimistic.

Scope

For: deciding whether to spend decode budget on a stream that has no real detail, media triage, dataset hygiene (finding upscaled images in a training corpus), and quality auditing on cheap hardware.

Not for:

• Not a deepfake or AI-generation detector. It detects resampling, nothing else.
• Not a general image-quality score. An upscaled image is not necessarily a bad one.
• Not provenance or authenticity. Resampling happens for many innocent reasons.
• Not a substitute for reading the container metadata when that is available and trustworthy.

Usage

import cv2, numpy as np, onnxruntime as ort
FACTORS = [2, 3, 4]
sess = ort.InferenceSession("upscale_factor.onnx", providers=["CPUExecutionProvider"])

img = cv2.imread("frame.png", cv2.IMREAD_GRAYSCALE)
p = img[y:y+64, x:x+64].astype(np.float32)        # a NATIVE 64x64 crop -- do not resize first
p = (p - p.mean()) / (p.std() + 1e-8)
print(FACTORS[int(sess.run(None, {"input": p[None,None]})[0][0].argmax())])

Take a native crop; never resize the image to 64×64 first. Resizing is itself a resampling operation and destroys the very signature the model reads. This is the single easiest way to get meaningless output.

Patches are cheap — run several across the frame and take the majority vote.

Known failure modes

1. Assumes the image WAS upscaled. The three-class head always answers 2×, 3×, or 4×. Gate it with the binary head or your own check first.
2. Trained on five interpolation kernels (nearest, linear, cubic, Lanczos4, area), randomised per example. A kernel outside that set — a neural upscaler, for instance — is untested and likely fails.
3. Grayscale, 64×64, single crop. Chroma upsampling artefacts are not used.
4. Heavy compression after upscaling will attenuate the periodic signature. Untested on strongly re-compressed material.
5. Flat regions carry no information. Sky, walls and blur have nothing to resample. Sample textured crops.

Training

• 2,600 COCO images, native 64×64 crops, split by source image (75/25)
• Downscale by the factor with INTER_AREA, restore with a randomly chosen kernel
• 4 conv layers (16→32→48→64), BatchNorm, global average pool
• Adam 3e-3, 22 epochs, batch 64

Verification

ONNX vs PyTorch, both CPU, 256 inputs: max relative logit difference 2.8e-07, 100% argmax agreement.

What "scalar baseline" means on this card

Every margin quoted here is against a stated baseline, because a margin without one is not a measurement. The baseline is the best single-threshold classifier over ten cheap statistics, fitted optimistically:

mean · std · lapvar · hf (high-frequency energy ratio) · grad (Sobel magnitude) · entropy · centre_edge · radial_slope · row_fft_peak · col_fft_peak

The last four are spatially aware, added after an earlier six-statistic baseline — all global aggregates — was found to systematically overstate model value on spatially structured tasks. A baseline that cannot see where anything is loses to a CNN by default. On one test task that flaw inflated an apparent margin from +0.060 to +0.261.

Two questions are asked with it, and they disagree:

• in-sample — threshold fitted on the data it is scored on. Deliberately generous. Answers is there structure beyond a low-order statistic?
• transferred — threshold fitted on the training corpus, applied unchanged to the target. Answers what should I ship? On one task the in-sample figure was 0.954 and the transferred figure 0.565.

Where this card quotes a single scalar figure without qualification, it is the in-sample one.

Provenance

COCO val2017, a public dataset. No personal data involved.