Which AI models does Airpic use, and under what licence?

Why it matters which model is behind the button

Most online photo tools describe what they do with a single word: enhance, magic, AI. The model doing the work is never named, so you cannot check three things that matter before you trust a result with a client's picture or a family photo.

Reproducibility. A named model with public weights behaves the same way every time. If a tool says it runs Real-ESRGAN x4plus, you know what to expect: real detail on photographic textures, a tendency to smooth fine grain, no new objects invented. If the tool is a black box, the behaviour can change overnight when the operator swaps the model, and the result you approved last month may not be the result you get today.

Licence. Open weights are not automatically free for business. Some of the best-known research models are published under non-commercial terms, which means a wedding photographer or a print shop cannot legally sell the output. When the model is named, you can read the licence yourself. When it is not, you are taking the operator's word for it.

Realistic expectations. Each model does one job. A restoration model works on the pixels of your file and adds coherent detail to them. A generative chat assistant does something different: it produces a new one- or two-megapixel image that resembles yours, with faces, text and proportions reinterpreted along the way. We explain the difference in detail in why an AI chatbot changes your picture when it enhances it. Knowing which kind of model you are using tells you in advance whether a damaged 60-pixel face can be recovered (it cannot, by any model) or whether a blurry 1200-pixel face can (usually, yes).

That is why we publish the list below, and keep it updated on the About Airpic page, which is linked in the footer of every page on the site.

Real-ESRGAN and GFPGAN online: the full list, tool by tool

These are the models running behind every tool on Airpic today, with the licence of each one. Nothing is left out and nothing is proprietary.

• Upscale an image: Real-ESRGAN, weights x4plus for photographs and x4plus-anime for flat artwork. Licence BSD-3-Clause. If you have been searching for Real-ESRGAN online, this is exactly that model, running on a GPU instead of your laptop. ×2 and ×4 are a single pass; ×8, ×16 and larger run the same model over tiles of the image, up to 600 megapixels in total. The anime weights have their own page, the anime and illustration upscaler, because they sharpen line art and flat colour in a way that looks wrong on skin and fabric.
• Remove the background: BiRefNet, licence MIT. The mask is computed at full resolution, so individual hairs and semi-transparent edges keep their shape without a halo. Output on transparent, white, black or any hexadecimal colour. Try it on the page to remove the background from a photo.
• Colourise a photo: DDColor, licence Apache-2.0, with a natural and a vivid variant and an intensity slider. Used on its own in the page to colourise a black-and-white photo and as the last stage of the old photo restorer.
• Remove noise: SCUNet, licence Apache-2.0, the real-noise PSNR weights, with an intensity slider so you can keep some grain on purpose.
• Unblur: Restormer, licence MIT, with separate weights for motion blur and defocus blur. The page to sharpen a blurry photo picks the variant for you after measuring the direction of the blur, and you can override it.
• Remove objects: LaMa (big-lama), licence Apache-2.0. You paint a mask in the browser and the model fills each region from its surroundings. It is the engine behind removing an object from a photo.
• Restore faces: GFPGAN v1.4, licence Apache-2.0, with our own face alignment and paste-back code. This is the GFPGAN online that people search for, with one important difference explained in the next section.
• Restore with reference photos: FLUX.2 klein 4B, licence Apache-2.0. When you add one to three other photos of the same person, this model generates face candidates guided by them, and a face-similarity check keeps only the candidates that still match the original expression.
• Transcribe video: Whisper large-v3, licence MIT, 99 languages, with sentence timing for the SRT and VTT files. Used by the page to transcribe a video to text and by the shorts maker.
• Summaries and chapters: Qwen3, licence Apache-2.0. It writes the summary, key points and chapters of the video to document tool and picks the moments for shorts. It never sees your video, only the transcript.

Compression is the one tool without a neural network: it is a plain encoder running on a CPU, which is why it takes seconds and is free to download without credits.

What a commercial licence means for you, and the models we turned down

In plain terms: every model above can be used to produce work that you sell. The BSD-3, MIT and Apache-2.0 licences all allow commercial use of the software and of what it produces. So a background removed for a shop listing, a colourised photo sold as a print, or a transcript delivered to a client are all covered on the model side. The copyright of the photo itself is a separate matter and stays with whoever owns the original; Airpic never claims any right over your files.

The reverse is also true, and it is the honest warning of this article. Some of the most impressive open models are not usable this way. There is a widely used face-restoration model whose weights are published under a non-commercial clause, and there are popular upscaling checkpoints whose licence is either non-commercial or simply unclear. They can give excellent results in a demo. We do not run them, because a tool that charges for a download cannot build on a model that forbids commercial use, and because you would be inheriting that restriction without knowing it.

GFPGAN is the case that needed the most work. The v1.4 weights and the network are Apache-2.0, but the helper code that detects a face, aligns it to a fixed template and pastes the restored face back into the photo came, in the reference implementation, from a research project with a non-commercial lineage. Rather than ship that code, we wrote our own: detection with RetinaFace, alignment from five facial landmarks, a soft mask from a face-parsing network, colour matching to the original skin tone and a multi-scale blend so there is no visible seam between the restored face and the rest of the picture. The restoration model is GFPGAN; everything around it is ours and was built with the licence question in mind. It also fixed a quality problem: the original paste-back left faces a shade darker than the body and sharp against a blurry neck.

A commercial licence does not mean the result is always right. Two things follow from how these models work. First, a restoration model guesses plausible detail; it does not know the person. On a badly compressed photo, pushing face fidelity to 100 % produces a clean but generic face that looks slightly like somebody else. In our tests, 40-60 % keeps the person recognisable. Second, running the face model several times does not add fidelity; it only stacks the model's own idea of a face. If you want the result to look more like the person, give it reference photos, not more passes.

Where each model runs and the limits of each tool

All of the models listed above run on GPUs in data centres in the European Union, and your file follows one path: it is uploaded, processed, offered to you as a preview, and deleted at most one hour after you download it, or one day if you never do. Nothing is used to train anything. The reasoning and the legal side are in our article on AI image processing in the EU.

Because the models run on real hardware with finite memory, every tool has a size ceiling. These are the current limits:

• Upscale: ×2 and ×4 as a direct pass on inputs up to 32 megapixels; ×8, ×16 and the maximum mode process the image in tiles up to 600 megapixels of output, which is the size you need for large-format print. The guide to upscaling an image for print explains how to choose the factor from the print size and viewing distance.
• Restore old photos: inputs up to 20 megapixels, output at ×1 or ×2. A typical flatbed scan at 600 ppi of a 10 × 15 cm print is about 2400 × 3600 px, which is 8.6 megapixels, comfortably inside the limit.
• Unblur: up to 25 megapixels, so a 48-megapixel phone photo needs to be downsized first.
• Remove the background, colourise, remove noise, remove objects: standard images up to 32 megapixels.
• Compress: runs on a CPU, not a GPU, which is why it costs nothing and finishes in seconds even for a 32-megapixel file.

Accepted input formats are the same everywhere: JPG, PNG (including alpha and 16-bit), WebP, TIFF, HEIC and AVIF. The EXIF orientation is applied and the ICC colour profile is kept, which matters if the file is going to a printer.

Two things about timing that a technical reader will want to know. Processing and previewing are free; you only spend a credit when you download the full file, and you get one full download free every month. And the first job after a quiet period can add 40-90 seconds while a GPU worker starts and loads the weights; once the worker is warm, the timing in the next section is what you should expect.

Open-source models for restoring old photos: the table and a step by step

The About Airpic page, linked in the footer, keeps the table of every model with its version, its licence and a link to the source repository, and it changes whenever a model changes. The why Airpic page sets out the same promise from the user's side: your pixels, no account, no copies.

Here is what those models do together in the tool to restore old photos, with the settings we recommend and why. The timing is a real measurement: on 19 September 2026 an old photo of 1000 × 1289 px was restored in 9 seconds on a warm GPU, in three stages (clean, faces, colour). The same day, a scan that had gone through the scanner 2.5 degrees askew was straightened automatically with a correction of 2.48 degrees.

1. Drop the scan. JPG, PNG, TIFF or HEIC, up to 20 megapixels. A file card shows the thumbnail, the name and the input and output sizes. Scan at 600 ppi if you still have the paper print; a phone photo of the print works but adds glare and keystone distortion.
2. Choose the output size. ×1 keeps the dimensions; ×2 doubles them. For a 1000 × 1289 px scan, ×2 gives 2000 × 2578 px, about 17 × 22 cm at 300 ppi. Behind this step is Real-ESRGAN x4plus with a reduction to the size you asked for.
3. Tick Straighten if the scan is tilted. The straightening only acts when it finds enough straight lines and the angle is at least a quarter of a degree; otherwise the photo comes back untouched with a note.
4. Tick Repair if there are scratches or tears. A mask editor appears; paint over the damage with the brush. LaMa fills those regions before anything else runs, so the face model never sees a crack across a cheek.
5. Tick Enhance faces and set the fidelity slider. It starts at 60 %. Leave it there for an ordinary scan; drop to 40-50 % if the source is a heavily compressed copy, because at 100 % GFPGAN invents a clean face that is no longer quite the person. If you have other photos of the same person, add up to three under Reference photos: the slider switches to 85 %, FLUX.2 klein 4B takes over, and you get several candidates with a similarity score to choose from.
6. Tick Colourise if you want colour. DDColor runs last, on the already restored image.
7. Press Process. The progress bar fills and stage chips appear one by one: Straighten, Damage, Clean, Faces, Colour. Click a chip to show the image as it was after that stage in the before-and-after comparator, and use View at 100 % to inspect eyes and teeth at real pixel size. If a stage went too far, untick it and process again; previews are free.
8. Download. PNG by default, JPG if you prefer. This is the only step that uses your free monthly download or a credit, and the file is deleted from our servers within an hour.

For the scanning side (resolution, dust, slides and negatives), read how to restore old photos with AI. For the models themselves, you now have the list, the licences and the limits, which is everything we would want to know before uploading a photo we care about.

FAQ

Keep reading