How to Restore Old Photos with AI — Complete Guide 2026

I found a box of my grandparents' photos in the attic. Wedding pictures from 1952. Faded to sepia. Scratched. One had a coffee stain from what I'm guessing was the Reagan administration.

I spent a weekend learning how AI photo restoration actually works. Here's what I learned, in order.

Step 1: Scan the Photo Properly

Skip this step and nothing else matters.

Use a flatbed scanner, not your phone. Set the resolution to at least 600 DPI. Save as TIFF or high-quality JPEG — never compress at this stage. A good scan captures more than you can see: details hidden in shadows, texture in faded areas, paper grain that the AI can use as reference.

If you only have a phone, use the highest resolution setting. Hold the phone parallel to the photo. Natural daylight, no flash. Take multiple shots and pick the sharpest.

Step 2: Assess What Can (and Can't) Be Fixed

AI restoration is good at: scratches, dust spots, fading, mild discoloration, low contrast, minor tears.

AI restoration struggles with: large missing sections (if half the face is torn off, the AI has to guess — and it will guess wrong), heavy water damage that warped the paper, photos that were folded and creased so badly the emulsion cracked.

Set realistic expectations. A moderately scratched photo from the 1970s will look nearly new. A photo that went through a flood and a fire? The AI will improve it, but it won't be a miracle.

Step 3: Run the Restoration

Our photo restorer gives you two paths:

Auto mode uses Topaz Dust & Scratch v2 — a model trained specifically on physical photo damage. It detects scratches, dust, and fading, then reconstructs the missing information. This handles 80% of restoration cases.

Face Pro mode adds GFPGAN on top. GFPGAN is a face restoration model — it identifies facial features and reconstructs them with realistic detail. Use this for portraits and group photos where faces are the priority. It's remarkable at restoring eyes, skin texture, and facial contours that were lost to fading.

Processing takes 10-30 seconds. Download the result and compare side by side with the original. The difference is usually dramatic for moderate damage.

Step 4: Add Color (If You Want)

Restoration fixes damage; colorization adds life. Our colorizer uses FLUX Kontext Pro — a model that understands context. Grass should be green. Sky should be blue. Skin tones vary by lighting and ethnicity.

Pick a style: Natural, Vibrant, Portrait, or Classic. Add an optional description to guide specific colors: "navy blue suit, red tie, golden afternoon light." The AI uses this alongside its own context understanding.

Best practice: restore first, colorize second. Fixing damage on a colorized image can produce weird artifacts where the color and the repair interact.

Step 5: Upscale for Archiving

After restoration and colorization, upscale the final result to 2x or 4x resolution. Choose Photo mode for real photographs. This step is optional but recommended if you want to print the restored photo or archive it digitally at high resolution.

My grandparents' wedding photo now hangs in my parents' hallway. It took about 5 minutes of actual work. The AI did the heavy lifting.