Old Photo Restoration Before and After: What AI Actually Does

Before-and-after examples of AI photo restoration can look dramatic — but expectations matter. Here's an honest look at what AI restoration actually changes, what types of photos see the biggest improvement, and what can't be recovered.

---

What Changes in the Before → After

AI restoration using CodeFormer + GFPGAN + Real-ESRGAN addresses three distinct problem categories:

1. Face Detail (CodeFormer)

Before: Faces that have softened and lost fine detail due to photographic paper aging. Eyebrows merge into forehead. Eyes become indistinct. Skin texture disappears into a flat blur.

After: Facial structure recovers. Eyes become defined. Facial features become identifiable. The subject is recognizable rather than approximate.

Best impact: Portraits from the 1940s–1970s where faces are the primary subject. A 1955 grandmother's portrait can go from "a person standing there" to "a specific person I recognize."

2. Fading, Yellowing, Color Shift (GFPGAN)

Before: White elements appear yellow or sepia-toned. The overall image looks flat and low-contrast. Colors have shifted to orange or amber.

After: Tonal range is restored. White is white. Contrast is restored. The image looks as it might have when it was taken, not as decades of photographic paper aging have left it.

Best impact: 1950s–1970s prints with systematic fading. 1970s–1980s color prints with orange color shift.

3. Resolution and Sharpness (Real-ESRGAN)

Before: The scan at 600 DPI produces a file that's usable but soft — printing at 8×10 introduces visible pixelation or softness.

After: The upscaled file prints cleanly at 8×10 and larger. Detail is synthesized to fill in resolution.

Best impact: Small original prints (2×3, 3×5) that need to be printed at larger sizes.

---

Photo Types by Expected Improvement

Maximum Improvement

Portrait, heavily faded, 1940s–1960s: This is where AI restoration provides the most dramatic change. Face reconstruction + fading correction + upscaling each make a significant contribution. Before: barely recognizable, washed-out portrait. After: clear, identifiable person with detail.

Black-and-white portrait with scratches: CodeFormer handles the face reconstruction, GFPGAN corrects the flat gray appearance, scratch removal targets physical damage. The cumulative improvement is substantial.

Good Improvement

Group photos, 1950s–1970s: Fading correction and upscaling are significant. Face reconstruction still applies but the smaller face size in a group photo reduces the per-face impact.

Color photo, 1970s–1980s: Color shift correction is the primary improvement. Faces in color photos from this era typically have better underlying detail than older B&W prints, so CodeFormer's work is less dramatic.

Moderate Improvement

1980s–1990s photos: Less degraded to begin with. Fading correction is the main improvement. Faces are usually clear enough that reconstruction has less to add.

Photos with major physical damage (large tears, heavy water staining): AI handles moderate damage very well. For large missing sections, the AI reduces the visual impact but cannot fully reconstruct lost content.

Limited Improvement

Very dark originals: Badly underexposed originals have limited underlying information. AI can brighten but can't recover detail that wasn't captured.

Very blurry originals: If the original photo was out of focus when taken, AI sharpening can help but camera motion blur or defocus isn't fully recoverable.

Extremely deteriorated prints: Photos that have deteriorated to the point where faces are essentially indistinct shapes have less to reconstruct from.

---

Managing Expectations

The output is a reconstruction, not a miracle. AI restoration works from the information present in the scan. Severely degraded photos see significant improvement, but the quality ceiling is determined by what's in the original.

Faces see the most dramatic improvement. Non-face content (backgrounds, objects, landscapes) improves through fading correction and upscaling, but the face reconstruction model is what produces the most visible change.

Quality of the scan matters. A 600 DPI scan gives the AI more information to work with than a 300 DPI scan. The before-after difference is better with a good input scan.

---

The 30-Day Guarantee

ArtImageHub offers a 30-day guarantee — if the restoration result isn't what you expected, you get a refund. For photos where results vary (severely damaged originals), this removes the risk from trying.

---

Restore your old family photos at ArtImageHub — $4.99 one-time →

Results in 30–90 seconds · HD download · 30-day guarantee

---

Related

• Photo Restoration Tips — how to get the best results
• How to Digitize Old Photos — scanning guide
• How to Restore Black and White Photos — B&W specific guide
• Best AI Tools for Old Photo Restoration in 2026 — 7-tool comparison