Restoring Old Photos with AI Upscaling

By ImgScale Team  ·  June 2026 · 6 min read · by kizura

Old family photos are one of the best use cases for AI upscaling. A 300×400 scan of a 1970s print can come out at 1200×1600 with recovered facial detail, sharpened hair, and restored texture in clothing — all in a couple of clicks. But getting the best results requires understanding what AI upscaling can and can't do, and preparing the source image correctly.

What AI upscaling does to old photos

AI upscaling (Real-ESRGAN) was trained on millions of image pairs — high-resolution originals degraded in various ways, alongside the degraded versions. This training means the model learned to reverse common degradation patterns: low resolution, JPEG compression, film grain, and print texture.

For old photos specifically, the model tends to:

• Recover facial features — eyes, eyebrows, mouth, and hair texture often come back clearly because the model has seen many faces during training
• Sharpen edges — the boundary between a person and a background, between hair strands, between fabric textures
• Reduce film grain — the model partially suppresses random grain as part of its restoration process
• Recover print texture — the halftone dots or silver crystal patterns of old prints are often smoothed into what the original scene looked like

What the model can't do: invent information that genuinely wasn't in the image. A severely out-of-focus face will be upscaled, but the face details won't appear — there was no information for the model to work with.

Scanning tips

If you're working from physical prints, the scan quality matters enormously. Tips:

• Scan at 600–1200 DPI minimum. 300 DPI prints barely give the AI enough to work with. 600 DPI gives much better results. 1200 DPI is ideal for small prints.
• Use a flatbed scanner, not a phone camera. Phone photos of prints introduce motion blur and uneven lighting that the AI treats as content.
• Clean the scanner glass. Dust spots become permanent features of the upscaled result.
• Scan in color even if the photo is black-and-white. A grayscale scan loses tonal information that helps the AI. Color mode captures the full range.
• Save as TIFF or PNG from the scanner, not JPEG. You'll be adding JPEG compression at export — don't add it at input too.

Pre-processing before upscaling

A clean source gives the AI the best possible starting point. Before upscaling:

Remove dust and scratches

Spot heal or clone stamp any visible dust spots, scratches, or tears. These will be "restored" as content by the AI — a scratch in the sky will come out as a sharp scratch in the upscaled output, not removed.

Correct exposure and contrast

If the scan is too dark or too flat, adjust brightness and contrast first. The AI upscales whatever it receives — a flat, low-contrast source produces a flat, low-contrast output. Bring out the tones before upscaling and the result will have more detail.

Reduce extreme noise

Heavy film grain or scanner noise can confuse the AI into treating it as high-frequency content (fine texture) rather than noise. A mild noise reduction pass before upscaling often helps. Be conservative — don't over-smooth, or you'll lose real detail too.

Don't sharpen before upscaling

Sharpening creates artificial high-frequency patterns that the AI will treat as real detail and amplify. Sharpen after upscaling instead.

Results by photo type

Portraits and group shots

Typically excellent results. The model is particularly good at faces. Individual faces larger than 50×50 px in the source almost always come out sharp with recognizable features in the upscaled result.

Outdoor landscapes

Very good. Natural textures (grass, trees, stone) upscale well because they have regular fine structure the model learned to reconstruct. Sky gradients come out smooth.

Interior scenes

Good. Architectural elements, furniture, and fabric textures benefit from the resolution increase. Watch for pattern amplification on complex textiles.

Old black-and-white photos

Excellent, especially for portraits. The model often produces striking results on B&W photos because the tonal structure maps cleanly to its learned patterns without the complexity of color matching.

Very old or damaged photos (pre-1920)

Variable. Extreme degradation, severe scratches, or large missing areas limit what the model can do. Expect improvement but not full restoration.

Maximizing face recovery

If the photo's main subject is a face and the person is small in the frame:

• Crop to the face first. Upscale the cropped version for better face recovery. A face that fills most of the frame gives the model more pixels to work with.
• The AI generally needs the face to be at least 40–50 pixels across in the source to produce a sharp result. Smaller faces may be upscaled but with less recognizable detail.
• After upscaling, use a face restoration tool (like GFPGAN or CodeFormer, available in other tools) for additional enhancement beyond what resolution recovery can do. PixelLab's upscaler is a general-purpose model, not face-specific.

Post-upscaling refinements

After downloading your upscaled result:

• Light sharpening: unsharp mask, radius 0.5–1px, amount 40–60%. Brings out edge crispness.
• Shadow noise check: zoom to 100% in a dark area. If noise is amplified, apply targeted noise reduction only to shadows.
• Color correction: old photos often have color casts (yellowing, fading). Correct after upscaling, not before.
• Save as PNG for the archive copy. Export a JPEG for sharing.

Honest limitations

AI upscaling is powerful but not magic:

• Motion blur is not removed — it's a different problem from low resolution
• Out-of-focus subjects are upscaled but not sharpened — the model can't invent detail that wasn't captured
• Very small faces (under 30 px) may be upscaled into a plausible but not accurate reconstruction
• Extreme JPEG compression at the source can produce artificial-looking "smoothed" skin or textures
• Fine text in old photos (names on signs, handwriting) may or may not be recoverable depending on source size

The best realistic expectation: your upscaled photo will look noticeably sharper and more detailed than the original, and faces will be more recognizable, but it will not look like it was taken yesterday on a modern camera. It will look like a well-restored older photo.

Related: Complete AI Upscaling Guide · PNG vs JPG vs WEBP Guide