Originally published on the HivePic blog. Disclosure: I work on HivePic, which is one of the four tools tested here.
Most "best unblur tool" posts compare screenshots by eye. We wanted a number, so we built a small test where the right answer is known: start from sharp photos, degrade them ourselves, run the degraded versions through web tools, and score each output against the original.
The headline result is a bit counter-intuitive. The tools made images sharper than the original, while moving them further from the original. That pattern held on the later HivePic runs too, including a dedicated Motion Blur mode that looked much better than the default.
1. Building a ground-truth set
We used four sharp photos we had generated ourselves (no people, no brands) and applied one degradation to each with Pillow. Each output stays pixel-aligned with its clean base.
| Sample | Degradation | Size |
|---|---|---|
| focus_soft_beach | Gaussian blur, radius 6 | 2304×1728 |
| motion_shake_dog | 31 px horizontal motion blur | 2560×1440 |
| lowres_upscaled_mug | 1/6 downscale → bicubic upscale | 2304×1728 |
| heavy_jpeg_cafe | 1/3 downscale → JPEG q12 → upscale | 1728×2304 |
Here is a compact version of that recipe:
from PIL import Image, ImageFilter
import io
import numpy as np
def out_of_focus(img): # sample 1
return img.filter(ImageFilter.GaussianBlur(radius=6))
def camera_shake(img, length=31): # sample 2: horizontal box kernel
a = np.asarray(img, dtype=float)
k = np.ones(length) / length
out = np.apply_along_axis(lambda r: np.convolve(r, k, mode="same"), 1, a)
return Image.fromarray(out.clip(0, 255).astype("uint8"))
def thumbnail(img, f=6): # sample 3
w, h = img.size
return img.resize((w // f, h // f), Image.BICUBIC).resize((w, h), Image.BICUBIC)
def chat_app(img, f=3, q=12): # sample 4
w, h = img.size
buf = io.BytesIO()
img.resize((w // f, h // f), Image.BICUBIC).save(buf, "JPEG", quality=q)
return Image.open(buf).resize((w, h), Image.BICUBIC)
2. The tools, and what came back
Fotor, PicWish and Pixelcut ran on 18 November 2026, in a desktop Chrome browser, with no trials and no new paid plans. HivePic was tested the same way, using the signed-in free daily allowance, and those runs were spread across days.
- HivePic (Super Focus V3, Standard mode, unless noted). The tool is built for the blur you actually have: Super Focus V3 is the default, and you can switch to a dedicated model such as Motion Blur. Sign in for 10 credits/day (5 per run, so two runs) and a full-resolution, same-size, watermark-free PNG (about 6 MB in our tests). A run takes about 32–40 s, plus 10–15 s to download, because it returns that full-resolution image rather than a downscaled preview — worth the short wait for a full-size file. Uploaded photos and results are automatically deleted after 7 days, so they don't sit on a server. Two minor limits, both stated on the site: the 10-credit daily allowance needs a Google sign-in, and signed-out downloads are Standard (1024 px on the long edge), while HD at original resolution is for signed-in users. Signed out it also gives only 3 credits/day, which is not enough for one unblur run. All of our HivePic runs used the signed-in free daily allowance; we spread the tests across days.
- Fotor. No login. It gave two free runs (about 15 s each), and the third image hit an upgrade modal. The free "Low-res" download was 2x the input (4608×3456, 5000×2812). Images 3–4 were not run.
- PicWish Photo Enhancer. All four processed without login in 10–14 s, but downloading needs a login and the free size is 1280 px on the long edge (per its Save menu). We didn't download.
- Pixelcut. Its unblur page is the "Fast Upscale" upscaler (2x/4x). All four results exceeded 4,000 px, which blocks the free download. We got zero files.
That leaves a measurable set of HivePic on all four images, plus a Motion Blur retest on the dog, and Fotor on the beach and the dog. PicWish and Pixelcut could only be judged from previews.
3. The metrics
Two cheap, well-known numbers:
- PSNR vs the clean base. This measures fidelity. Higher means the pixels are closer to what was actually there.
- Variance of the Laplacian. This is a no-reference sharpness score that measures how much edge energy an image has. Higher means crisper. It says nothing about whether those edges are correct.
Each output was converted to grayscale and resized (Lanczos) to the base's size. Fotor's 2x outputs were scaled down, and HivePic's were already the same size.
from PIL import Image
import numpy as np
def load_gray(path, size=None):
img = Image.open(path).convert("L")
if size:
img = img.resize(size, Image.LANCZOS)
return np.asarray(img, dtype=float)
def psnr(x, ref):
mse = ((x - ref) ** 2).mean()
return 10 * np.log10(255 ** 2 / mse)
def laplacian_var(a):
lap = (4 * a[1:-1, 1:-1] - a[:-2, 1:-1] - a[2:, 1:-1]
- a[1:-1, :-2] - a[1:-1, 2:])
return lap.var()
ref = Image.open("focus_soft_beach_base.jpg")
base = load_gray("focus_soft_beach_base.jpg")
for name in ["focus_soft_beach_blurred.jpg", "hivepic_1_beach.png", "fotor_1_beach.jpg"]:
x = load_gray(name, size=ref.size)
print(f"{name}: PSNR {psnr(x, base):.2f} dB, Laplacian var {laplacian_var(x):.1f}")
print(f"original: Laplacian var {laplacian_var(base):.1f}")
4. Results
| PSNR (dB) | Laplacian var | |
|---|---|---|
| Beach: blurred input | 26.68 | 1.2 |
| Beach: HivePic | 24.47 | 358.5 |
| Beach: Fotor | 25.79 | 221.2 |
| Beach: clean original | n/a | 51.0 |
| Dog: blurred input | 21.18 | 11.1 |
| Dog: HivePic default | 20.45 | 223.4 |
| Dog: HivePic Motion Blur | 20.03 | 526.2 |
| Dog: Fotor | 19.82 | 316.3 |
| Dog: clean original | n/a | 84.2 |
| Mug: blurred input | 32.06 | 2.2 |
| Mug: HivePic | 30.25 | 130.9 |
| Mug: clean original | n/a | 26.8 |
| Café: blurred input | 28.69 | 7.3 |
| Café: HivePic | 27.38 | 181.4 |
| Café: clean original | n/a | 58.7 |
Three things stand out.
Sharpness overshoots the original. On every measured output, Laplacian variance landed well above the real photo, not near it. A deblurring filter that only recovered what was lost would land near the original's score, not far past it. HivePic's Motion Blur result is the extreme case: 526.2, versus 223.4 in default mode and 84.2 on the original.
Fidelity goes down, not up. Every output has a lower PSNR than the blurry input it was given. Pixel for pixel, the "fixed" images are further from the truth than the blurred ones. That's what you expect from generative restoration: the model synthesizes plausible high-frequency detail (rock texture, ripples, fur, petals) instead of inverting the blur. The mug went from 32.06 dB to 30.25; the café from 28.69 to 27.38. The Motion Blur dog, which looks much sharper than the default, is slightly further from the original than the default was (20.03 dB vs 20.45, against an input of 21.18).
Visual quality and fidelity can diverge. On the mug, HivePic was moderately crisper, with a tighter rim, handle and linen weave and no artifacts we could see — a modest, clean lift that still didn't restore the original's fine ceramic speckle. On the café, the change was the biggest of the set: the crushed JPEG became a sharp, attractive scene, but details like the flower petals look reconstructed rather than recovered. The bouquet is convincing; it isn't the same bouquet. On the dog, default Super Focus V3 left the face soft and the 31 px smear obvious (same story for Fotor). Motion Blur mode largely removed the smear and sharpened the face clearly, with faint horizontal streaks and a slightly painterly texture left behind. We did not test a dedicated motion-blur mode on Fotor, PicWish or Pixelcut, and we have no Fotor file for the mug or the café (its free allowance ended on image 3). PicWish and Pixelcut remain preview-only.
The visual sheets match those notes.

Beach and dog, default models. Left to right: input, HivePic, Fotor, original.

Mug and café: input, HivePic, original. Dog: input, HivePic default, HivePic Motion Blur, original.
5. Caveats
- Small set: metrics for HivePic on four images plus one Motion Blur retest, and for Fotor on two images. Treat this as a demonstration of the pattern, not a ranking.
- PSNR on grayscale after resampling is crude. SSIM or LPIPS would add nuance, and LPIPS would probably be kinder to generative outputs.
- We tested HivePic's dedicated Motion Blur mode only on the dog, and only against the other tools' default modes.
- Fotor stopped at its free limit on image 3, so there are no Fotor numbers for the mug or the café. PicWish and Pixelcut have no downloadable files in this test, so they have no metrics.
- Free tiers, prices and limits are as observed on 18 Nov 2026.
Takeaway
If you build or pick an unblur tool, measure fidelity and sharpness separately. Sharpness alone rewards hallucination. For holiday photos that trade-off is fine. For plates, documents, or anything evidentiary, an "unblurred" image is a plausible guess, not a recovery. A dedicated motion-blur mode can look much better than a default enhance and still move the file further from the original. If you want to reproduce this with your own photos, the snippets above are all you need. If you want to try the tool I work on, it's HivePic's unblur tool (two free runs/day after sign-in).
Top comments (0)