AI photo enhancer: sharpness, size and old photos
An AI photo enhancer makes a photo larger without blocky pixels, removes noise and blur, and restores old pictures. For exact upscaling, use Topaz Labs or the free open-source Upscayl; for detail in small images, Magnific and Krea; for restoring by description, Nano Banana and image generation in ChatGPT. Remember: any such upscaling invents detail.
In short
- Upscale and sharpen without invented detail: Topaz Labs
- Free and on your own computer: Upscayl
- Add detail to a small image: Magnific, Krea
- Editing together with Photoshop: Adobe Firefly
- Restoring an old photo by description: Nano Banana, image generation in ChatGPT
The list
The order follows the criteria of this page; it is not a rating.
- #1
Topaz Labs
Topaz Labs, USA
- What it is good for
- Upscaling, sharpening and noise removal that treats the original details with care, for photos and video.
- Why it is on the list
- In September 2026, Topaz Labs released the Starlight Fast 3 model: according to the company, it upscales video at up to 7 frames per second at 1080p and runs 4 times faster than Starlight Precise. Topaz Labs, “Speed Update” for September 2026, 28 September 2026
- Price
- no free plan; paid from $19 per month (web version, Image plan) (16 October 2026) plans
- Russian language
- no per the site
- Official site
- topazlabs.com
- #2
Upscayl
open-source project, open-source project
- What it is good for
- Free photo upscaling on your own computer, with a Russian interface, and photos never go online.
- Why it is on the list
- As of 16 October 2026, the Upscayl repository on GitHub has 50.6 thousand stars. GitHub, upscayl/upscayl, 16 October 2026
- Price
- free (16 October 2026) plans
- Russian language
- yes per the site
- Access from Russia
- open model, downloadable with no country restrictions official page
- License
- AGPL-3.0
- Official site
- upscayl.org
- #3
Magnific
Magnific (formerly Freepik), legal entity Freepik Company S.L.U., Spain
- What it is good for
- Upscaling with added detail: the texture of skin, fabric and leaves in small images.
- Why it is on the list
- On 28 April 2026, Freepik changed its name to Magnific and brought its AI tools together under it; according to Tech.eu, the company has more than 1 million paying subscribers and $200 million in annual subscription revenue. Tech.eu, 28 April 2026
- Price
- free plan with limits; paid from £10.5 per month billed annually (16 October 2026) plans
- Official site
- magnific.com
- #4
Krea
Krea.ai, Inc., USA
- What it is good for
- Upscaling and editing in 2K by description and reference images, with a free daily limit.
- Why it is on the list
- On 15 October 2026, Krea released the Krea 3 model: it edits images at 2K resolution, accepts up to 6 reference images and costs $0.06 per 2K image through the API. Krea blog, 15 October 2026
- Price
- free plan: 100 units per day; paid from $35 per month ($21 per month billed annually) (16 October 2026) plans
- Official site
- krea.ai
- #5
Adobe Firefly
Adobe, USA
- What it is good for
- Extending the frame, removing unwanted objects and retouching, with further work in Photoshop.
- Why it is on the list
- Adobe states that Firefly models are trained on licensed Adobe Stock content and public domain material, and results carry Content Credentials, information about the use of AI. Adobe, Adobe Firefly page, 16 October 2026
- Price
- free daily generations; paid from $9.99 per month (Firefly Standard, 2000 credits) (16 October 2026) plans
- Official site
- firefly.adobe.com
- #6
Nano Banana (Gemini)
Google, USA
- What it is good for
- Restoring an old photo with a phrase: remove scratches, bring back color, redraw a lost corner.
- Why it is on the list
- On 6 October 2026, Google made Nano Banana 2.1 generally available, an update to Nano Banana 2, released in February 2026. Google AI for Developers, Gemini API changelog, 6 October 2026
- Price
- free in the Gemini app with limits; paid Google AI plans from £4.49 per month (the page showed prices in pounds) (16 October 2026) plans
- Russian language
- yes per the site
- Access from Russia
- not officially supported official page
- Official site
- gemini.google.com
- Glossary term
- Nano Banana (Gemini)
- #7
Image generation in ChatGPT
OpenAI, USA
- What it is good for
- Restoration and colorization by a detailed description, with several options to choose from.
- Why it is on the list
- On 21 April 2026, OpenAI released the gpt-image-2 model for ChatGPT and the API: it makes images up to 2K and up to eight consistent pictures from one prompt. IT Brief, 21 April 2026
- Price
- in the API, the gpt-image-2 model: $5 per 1 million input text tokens, $30 per 1 million output image tokens; ChatGPT plan prices not checked (16 October 2026) plans
- Access from Russia
- not officially supported official page
- Official site
- chatgpt.com
In short: which tool for which problem
The order on this page follows criteria, not a ranking. First come the tools that upscaleAI upscaling. AI upscaling raises the resolution of an image or a video with a neural network that does not just stretch pixels but draws in believable small details. a photo and keep the original details, then the services that invent detail. We found no open blind ranking for photo upscaling like the Arena boards for image generationtext-to-image. Text-to-image AI is image generation from a text description: you write what you want to see, and a neural network draws the picture., so there are no place numbers.
The choice depends on the problem. A small photo for print is upscaled with Topaz Labs or Upscayl. Topaz also removes noise from a night shot. A blurred face on an old print is restored by generative modelsgenerative AI. Generative AI is a family of neural networks that create new content: text, images, music, video and program code, from a description in plain language., but they invent features: the last section of the page covers this.
Upscaling and sharpness
Enlarging, or upscaling, means increasing the number of pixels in a photo. A regular editor stretches the pixels, and the image gets blurry. A neural network guesses what the new pixels should be from the photos it learned on, so the edges stay sharp.
How far you can upscale depends on the source. A phone photo taken in good light will hold up as a print on a poster, while an avatar from a messenger will not: it holds too little information, and the model will invent almost everything. Noise and motion blur are removed in separate modes, turned on before upscaling, not after.
Topaz Labs from the US makes software for photo and video quality. There is no free plan, the web version with the Image plan costs from USD 19 a month, and there is no Russian interface. The company is also developing video: in September 2026 it released the Starlight Fast 3 model, which, according to Topaz, works 4 times faster than Starlight Precise and processes up to 7 frames per second at 1080p. Video is covered in more detail in the list of AI tools for video editing.
Magnific (formerly Freepik, Spain) upscales with added detail: a small image becomes a large one with the texture of fabric and leaves. There is a limited free plan, and paid plans start at GBP 10.5 a month with yearly billing. On 28 April 2026 Freepik changed its name to Magnific and brought its AI tools together under it.
Krea gives 100 units a day for free, and the paid plan starts at USD 35 a month. The Krea 3 model came out on 15 October 2026; it edits images in 2K and accepts up to 6 reference images in one request.
Old and damaged photos
An old photo has every problem at once: cracks, stains, fading, low resolution after scanning. Generative models take on such a task from words: “remove the scratches, bring back the color, do not change the faces”.
Nano Banana in the Gemini app does this for free within limits, and on 6 October 2026 Google made version 2.1 generally available. Image generation in ChatGPT on the gpt-image-2 model gives up to eight consistent options per request, which makes it easy to pick the one closest to the original. Neither service officially works in Russia: it is not on ChatGPT’s list of countries, and Google does not include Russia in the list of regions for the Gemini API and Google AI Studio. We do not give instructions for getting around these restrictions.
Adobe Firefly fits if the retouching then goes to Photoshop: extend a cropped frame, remove unwanted objects, replace the background. There are free daily generations, and the paid plan starts at USD 9.99 a month for 2,000 credits.
Colorizing a black-and-white photo is also a guess: the model chooses the color of a dress, wallpaper or car by probability, not by fact. If the color is known from family history or an archive, name it directly in the description.
First scan the photo at the scanner’s maximum resolution, then restore it. Compare the result with other photos of the same person: the model may make a face sharper but someone else’s.
Open-source programs for your computer
Upscayl: a free open-source program for Windows, macOS and Linux, with a Russian interface. Photos never go online, and your graphics card does all the computing. As of 16 October 2026 the Upscayl repository on GitHub has 50,600 stars.
Working with Upscayl takes three steps: choose a photo or a folder, choose the model and the scale, press the button. The program processes a folder of hundreds of photos in one batch, which is handy for a catalog archive. Speed depends on the graphics card: on a laptop without a separate graphics card, processing is noticeably slower.
The AGPL-3.0 license in plain words: you can use the program for free, including in a company’s work. If you change its code and give people access to the changed version over a network, you will have to publish the source code of your changes. For processing your own photos this condition changes nothing.
Open generative models, such as Qwen-Image or FLUX, can also run on your own machine, but for restoration they need setup, which is a developer’s job. This path suits a company that must not send photos to a foreign cloud; it is covered in detail in the section “Sovereign AI”.
Where AI adds what was not there
Any AI upscaling invents details that were not in the original photo. For a postcard this is no problem, but for evidence it is: a car number plate, a face on camera footage or a signature on a document show after processing not what was there but what the model thought likely.
So an enhanced photo is not fit for documents, expert reviews or court. Keep the original separately and labeltarget variable. A target variable is the value a machine learning model learns to predict: whether a customer will leave, how much product will sell, how much a flat costs. the processed copy.
Generative services invent the most: Magnific openly sells added detail, and Nano Banana and ChatGPT may redraw a frame more than you asked. If you need edits without invention, start with Topaz Labs or Upscayl and compare the result with the original at 100% scale.
If you need to change a photo rather than restore it (replace the background, remove a person, change the light), see the list of AI for photos.
How we built this top
- How well the tool keeps the original details instead of inventing them
- Works on your own computer without sending photos online
- Price and free limit from the official page
- Russian language and access from Russia
The order follows the criteria of this page; it is not a rating.
Checked:
How we build ratings