Home › Methodology

How we test photo restoration

Last verified: 31 July 2026

This page is the standard behind every claim on this site. We restore real damaged photographs, sourced from public archives, through the same pipeline you use, and we publish the results unedited, including the failures and the cases where the model invents. When we compare tools, the same photos and rules apply.

Our bias, stated plainly

PicFix is our product. Anything on this site that judges restoration quality involves something we sell, and you should read it with that in mind. Two commitments keep us honest. Every test result is published, with the embarrassing ones given equal billing. And if you find an error anywhere on this site, an outdated price, a claim our own evidence contradicts, tell us through the support page and we will correct it and note the correction.

The test set

Our corpus is built from photographs with real damage, never staged: prints recovered after Hurricane Katrina, chemically faded slides from government archives, scratched and torn portraits spanning the 1850s to the 1990s. All of them come from Wikimedia Commons damage collections, under public domain or Creative Commons terms, credited beneath each exhibit. Each one goes through the production PicFix pipeline at default settings, one pass, and the outcome gets published whatever it is. Nothing here was rerun until it looked good.

How we grade

Grade	Meaning
Recovered	The picture survived in the print and came back faithfully: same people, same framing, same scene.
Partly reconstructed	Repair plus plausible fill where pieces were missing. Fine for backgrounds; check any face near the damage.
Invented	Too little survived, so the output imagines a picture. It can look excellent while showing people who were never there.
Drift	The framing or the content changed: a tighter crop, an altered pose. We count this as a failure even when the result is handsome.
Blocked	The safety checker refused the photo. Happens occasionally with archival material.

Grading is a per-photo review of the pair at identical sizes, which is also how every comparison on this site is displayed. A zoomed or cropped "after" can smuggle in flattery; equal pixels keep the comparison honest.

What recovery looks like

1972 color slide faded almost entirely to red — Grade: Recovered. A 1972 EPA slide with its dyes collapsed to red. The scene came back with the warning sign's lettering intact. Public domain (US National Archives), via Wikimedia Commons.

The same slide restored: blue sky, readable warning sign, natural color — Grade: Recovered. A 1972 EPA slide with its dyes collapsed to red. The scene came back with the warning sign's lettering intact. Public domain (US National Archives), via Wikimedia Commons.

Small faded sepia portrait of a young man in uniform, 1925 — Grade: Recovered. A 1925 portrait, faded and yellowed. The output stayed monochrome as instructed and the face matches the print. Public domain, via Wikimedia Commons.

The same portrait cleaned and sharpened, still monochrome — Grade: Recovered. A 1925 portrait, faded and yellowed. The output stayed monochrome as instructed and the face matches the print. Public domain, via Wikimedia Commons.

Where restoration ends

Restoration works with what survives. When almost nothing survives, a model this capable does something you should know about before you trust any tool, ours included: it fills the frame with an imagined picture.

Album pages caked in silt after Katrina floodwater, images destroyed — Grade: Invented. Pages from a family album after weeks under Katrina floodwater. Almost no image survived, so the model imagined people onto the pages. We would tell you the same if it were your album. Damaged album photographed by Infrogmation of New Orleans, CC BY 2.5, via Wikimedia Commons.

The output shows clear photographs of people on the album pages — Grade: Invented. Pages from a family album after weeks under Katrina floodwater. Almost no image survived, so the model imagined people onto the pages. We would tell you the same if it were your album. Damaged album photographed by Infrogmation of New Orleans, CC BY 2.5, via Wikimedia Commons.

Very small scratched portrait of a bearded man, around 1900 — Grade: Partly reconstructed. A print smaller than a matchbook, upscaled into a detailed portrait. The fine detail had to be synthesized, because a print this small never recorded pores or single hairs.

A detailed large portrait produced from the tiny original — Grade: Partly reconstructed. A print smaller than a matchbook, upscaled into a detailed portrait. The fine detail had to be synthesized, because a print this small never recorded pores or single hairs.

Cracked 1940s Navy training photo of a gymnast on a wheel apparatus — Grade: Drift. A cracked 1940s US Navy training photo. The crack is repaired, and the gymnast's grip and pose changed in the regeneration. We count this as a failure. Public domain (US Navy), via Wikimedia Commons.

The repaired photo, where the gymnast's grip and pose differ from the original — Grade: Drift. A cracked 1940s US Navy training photo. The crack is repaired, and the gymnast's grip and pose changed in the regeneration. We count this as a failure. Public domain (US Navy), via Wikimedia Commons.

Known limits today

Total loss produces invention. Where emulsion is gone, every tool on the market guesses; ours guesses in high resolution. Keep the original and treat such output as an artist's impression.
Severe color fade can read as monochrome, and the output comes back sepia. Rerunning usually resolves it; we are also tightening the color-mode handling.
Framing and content drift happens occasionally. We are building automatic drift detection that retries the job before you ever see it.
The safety checker sometimes refuses archival photographs it misreads. A different crop of the same print usually passes.
Your original file is never altered, and every result arrives next to the original with a slider, so you can run this same review on your own photo.

The standard for comparisons

When this site publishes tool-against-tool comparisons, they will follow the rules above with no exceptions: the same damaged photos through every tool, default settings, prices paid out of pocket, results shown at identical sizes, and the misses published next to the hits. Until a comparison meets that bar, we do not publish a verdict.

Corrections

Found something wrong here? Tell us through the support page. We correct the page, date the correction, and re-run affected tests when the pipeline changes in any way that could alter a published result.

Run your own test In your browser

No download. Photos are deleted from our servers after processing.