How to Restore Old Photos with AI — A Practical Guide

Most families have a shoebox of old photos somewhere — wedding pictures from the 1960s, military portraits, school photos with creases down the middle. Until recently, restoring these meant either paying a professional $50-200 per image or spending hours in Photoshop with the clone stamp tool.

AI restoration has changed the math. Tools like GFPGAN (Tencent ARC, 2021), CodeFormer (S-Lab, 2022), and Real-ESRGAN (Xintao Wang et al.) can process a damaged photo in 5-15 seconds. The quality gap between AI and manual restoration has narrowed to the point where, for most consumer use cases, AI output is good enough.

Step 1: Scan Properly

Garbage in, garbage out. A phone photo of a photo adds glare, distortion, and perspective warping that no AI can fully fix. If the original photo matters to you, scan it.

• Flatbed scanner: 600 DPI minimum for prints. 1200 DPI if you plan to crop or enlarge. Epson Perfection V600 (~$230) is the standard recommendation for home use.
• Phone scan app: If no scanner, use Google PhotoScan or Adobe Scan. These take multiple shots and stitch to remove glare. Not as good as a flatbed, but far better than a single phone photo.
• File format: Save as PNG or TIFF. JPEG adds compression artifacts that confuse AI models. If you must use JPEG, set quality to 95+.

Step 2: Identify the Damage Type

Not all damage is the same, and matching the repair mode to the actual problem saves credits and produces better results.

Step 3: Restore and Compare

Run the restoration and compare side-by-side. A few things to keep in mind:

• AI is probabilistic. Face enhancement models sometimes "hallucinate" details — they reconstruct what a face probably looked like, not what it actually looked like. For genealogical accuracy, keep the original alongside.
• Colorization is a guess. No AI model knows the actual colors of a B&W photo. DeOldify tends toward warm, saturated tones. DDColor is more neutral. Neither is "right."
• Multiple passes help. Try scratch repair first, then face enhance on the result. Chaining modes often beats "Full Restore" for specific damage patterns.

What AI Can't Fix

Some damage is beyond current AI capabilities:

• Torn photos with missing pieces. AI can fill small gaps but cannot reconstruct a missing half of a face.
• Waterlogged photos. Severe water damage creates patterns that AI models interpret as texture, producing strange artifacts.
• Overexposed/underexposed areas. Pure white or pure black regions have no data to recover. AI can guess, but results are often inaccurate.

Cost Comparison

Professional restoration: $50-200 per photo, 1-2 week turnaround. Adobe Photoshop subscription: $22.99/month, plus 2-4 hours of your time per photo if you learn the tools.

AI tools: $0.25-0.50 per photo, under 30 seconds. The tradeoff is control — a skilled Photoshop artist can produce better results on difficult cases. But for the 80% of old photos with common damage (scratches, fading, low resolution), AI output is comparable to what most people could achieve manually after hours of work.