Restore Old Photos with AI

Real results from Photo Restoration

A real result from this tool. Drag the handle to compare.

After

Before

Drag to Compare Before & After

Will Photo Restoration work on your photo?

Match what is wrong with your image to what this tool can realistically do. Where it cannot help, the row says which tool can.

• Scratches, tears, creases, stains

  Good match

  This is the core case. The instruction tells the model to repair the damage and leave faces, expressions and composition exactly as they are.

• Fading and yellowing

  Good match

  Pick the Vintage preset. Exposure and contrast are recovered without repainting the scene.

• A black and white original you want in colour

  Good match

  Turn on Add color. Colours are inferred from context, so treat them as plausible rather than documented.

• A missing corner or a small hole

  Partial improvement

  Small gaps are filled convincingly. A missing face cannot be recovered truthfully — whatever appears there is a guess.

• The print is intact but the face is just soft

  Use another tool

  There is no damage to repair. Sharpen the face instead with Unblur Face

How to restore old photos online

Old photo restoration comes down to three things: a good digital copy of the print, the repair that matches the damage, and a careful look at the result. Everything below runs in the browser, on a computer or a phone; there is nothing to install.

• Get the photo off the paper first

  A flatbed scan at 300 dpi or more gives the model the most to work with. No scanner? Photograph the print in daylight, phone held parallel to it, no flash, no glare, and take it out of the album sleeve first: plastic adds reflections the model will try to keep. Crop to the photo before uploading.

• Match the preset to what is actually wrong

  Scratches & tears is for physical damage: cracks, creases, missing specks. Vintage / faded is for yellowing and lost contrast. Black & white with Add color on is for colorizing. General old photo handles a mix of everything and is the safe default when you are not sure.

• Restoring a whole box of old photos

  Each photo is one run. Sort the prints by damage type before you scan so the preset stays the same for a batch, then work through them; every result is saved to My Creations, so you can download them all at the end instead of one by one.

• When to restore old photos by hand instead

  Manual retouching in Photoshop or a professional restorer is still the better choice for a single heirloom with a large missing area, especially a missing face. AI fills a gap with something plausible, not with what was there. For ordinary wear, fading and scratches, a 25-second run gets most of the way.

• Keep the original scan

  Save the unretouched scan next to the restored copy. Restoration is an interpretation, and colorization doubly so; for family history, the original is the record and the restored version is the one you share.

• Compared with Firefly, Fotor, Picsart and Krea

  Adobe Firefly, Fotor, Picsart and Krea put restoration inside a full editor or a broad AI suite. This page is only restore old photo AI: upload, pick the damage, optional colorize, download. Preview with the credits you get at sign-in; free-plan downloads carry a small watermark. Identity stays locked — the model is told not to beautify or reinvent faces.

Why Photo Restoration on AI Image Tools

• Repairs, doesn't repaint

  The instruction sent to the model insists on keeping every face, expression and object exactly where it is. It fixes the damage instead of imagining a new photo that looks similar.

• Tell it what is wrong

  Pick the damage type: general wear, vintage fading and yellowing, scratches and tears, or a black and white original. The prompt adapts to it.

• Color when you want it

  Leave the photo as it is, or flip Add color for natural skin tones and period-appropriate colors. Colorization is an interpretation, and it is off by default.

• A focused page, not a suite

  No Firefly canvas, Fotor workspace, Picsart project or Krea playground to learn. One restore tool: preview with signup credits, free downloads watermarked, then Unblur Face / Enhancer / Upscaler if you need the next step.

• A workflow, not one button

  Restore first, then sharpen faces with Unblur Face, clean up with the Enhancer and double the size with the Upscaler. The related tools are linked at the bottom of this page.

Every image model in one place

Pick any of these from the Model menu above; the price is shown before you generate. Each name links to what that model is good at. Compare every tier.

• Z-Image Turbo

  Alibaba · cheapest and fastest, text-to-image only

  1 credits per image

• Seedream 5 Pro · 1K

  ByteDance · same model as 2K at half the price

  9 credits per image

• GPT Image 2.5 Flare · 2K

  OpenAI · same price as Sunburst, tuned for speed and volume

  13 credits per image

• Qwen Image 3 · 2K

  Alibaba · strong at readable text inside the image

  6 credits per image

• Nano Banana

  Google Gemini 2.5 Flash · fastest, cheapest

  5 credits per image

• Nano Banana 2.1 · 1K

  Google · newest Nano Banana, sharper text, banner ratios

  10 credits per image

• Nano Banana 2.1 · 2K

  Google · same model at 2K

  15 credits per image

• Nano Banana 2.1 · 4K

  Google · for print or heavy cropping

  23 credits per image

• Ideogram V3 · Turbo

  Ideogram · text inside the image, fastest tier

  5 credits per image

• Ideogram V3 · Balanced

  Ideogram · text inside the image, posters and logos

  9 credits per image

• Ideogram V3 · Quality

  Ideogram · text inside the image, highest detail

  13 credits per image

• FLUX.1 Kontext

  Black Forest Labs · fastest, about 5s per image

  5 credits per image

• Seedream 5 Lite

  ByteDance · good at Chinese text and posters

  7 credits per image

• Seedream 5 Pro · 2K

  ByteDance · most photorealistic, highest detail

  18 credits per image

• GPT Image 2 · 2K

  OpenAI · good with text in images, half the price of 2.5

  7 credits per image

• GPT Image 2 · 4K

  OpenAI · for print or heavy cropping

  10 credits per image

• GPT Image 2.5 · 2K

  OpenAI · newest, better instruction following

  13 credits per image

• GPT Image 2.5 · 4K

  OpenAI · newest, highest resolution

  20 credits per image

Photo Restoration specifications

Input

One JPG, PNG, WebP or HEIC up to 10 MB; scan prints at 300 dpi or more

Damage presets

General old photo, Vintage / faded, Scratches & tears, Black & white

Colorization

Off by default; one switch to add natural color

Models

FLUX.1 Kontext by default; Nano Banana, Nano Banana 2.1, Seedream 5 Lite / Pro, GPT Image 2 / 2.5 and Ideogram V3 selectable

Sizes

Same as input, 1:1, 4:3, 3:4, 16:9

Output

PNG or JPEG depending on the model, saved to My Creations

Cost

5 to 20 credits per image, depending on the model

Time

About 25 seconds with FLUX.1 Kontext; Seedream around 50 seconds, GPT Image about a minute

Free plan

10 credits on sign-in — 10 images on the default model; free downloads carry a small watermark

How to use Photo Restoration

1. Step 1

   Scan or photograph the print

   Flat, evenly lit, no glare. A phone photo of the print works; a flatbed scan works better.

2. Step 2

   Pick the damage type and color option

   Choose the preset that matches the print and decide whether to colorize. The “Try a sample photo” button in the input box loads one if you want to test first.

3. Step 3

   Restore, then refine

   Check faces and edges at full size. If a face is still soft, send the result to Unblur Face; if you need a print, run the Upscaler.

Get a better result

A few choices before you upload change the output more than any setting on this page.

• Crop to the photograph itself

  Leave out the album page, the frame and the desk. Everything you include is detail the model spends attention on instead of the picture.

• Run it twice if you want colour

  Restore once with Add color off, then again with it on. You get both versions for the price of two runs and can pick per photo.

• Sharpen the faces afterwards

  Restoration fixes the paper; it does not target faces. For portraits, send the result on. Unblur Face