How to Restore Old Family Photos Locally (Without Uploading Them)
Old photo restoration is not a button, and the tools that market it as one are usually cloud services that want your family photographs on their servers. It is a pipeline of six stages — and the single most common mistake happens at stage one, before any AI is involved at all.
Six stages, and what each one needs
| Stage | What it fixes | What handles it |
|---|---|---|
| Scan | Capturing the most detail the print or negative still holds | Any scanner, ideally with VueScan or SilverFast |
| Denoise | Film grain, scanner noise, JPEG artefacts | Formatif AI — Denoiser, locally |
| Upscale | Resolution for printing or large screens | Formatif AI — AI Upscaler 2× / 4× |
| Face restore | Softness and degradation on the faces themselves | Formatif AI — Face Restorer |
| Scratches and dust | Physical damage, tears, spots, creases | Photoshop healing tools, or GIMP with Resynthesizer |
| Colourise | Black-and-white photographs | Photoshop Neural Filter Colorize, DDColor or DeOldify |
Four of these six stages happen without any upload. Two do not, and this page names every one of them rather than pretending otherwise.
The short answer
Scan at 600–1200 dpi for prints or 2400–4800 dpi for film, as 48-bit TIFF with all automatic processing switched off. Then denoise, upscale and restore faces — all of which run locally. Scratches need healing tools and black-and-white needs a separate colourisation model, so no single application covers all six stages.
That last sentence is the whole point of this page. If you are looking for one purchase that restores a shoebox of photographs end to end, it does not exist — locally or in the cloud. What you can do is get four stages done on your own machine, in one pass, without your family photographs leaving your computer.
Note on names: Formatif AI (formatif.net) is offline desktop media software, unrelated to Formative, the education platform.
The six-stage pipeline, and why the order matters
Restoration fails most often because stages are applied in the wrong order — upscaling before denoising, or colourising a faded scan. The sequence below is the one that holds up on real family archives:
- Scan at the correct resolution as an uncompressed master. Everything downstream is capped by what this step captures.
- Correct tone and colour while the image is at full resolution. Faded prints and colour casts respond well to levels, curves and white balance, and fixing them first means later stages are not working around a bad image.
- Repair physical damage — scratches, dust, creases, torn corners — with healing and inpainting tools. Do this before AI cleanup so you are not asking a denoiser to interpret a scratch as texture.
- Denoise the film grain and scanner noise that remain.
- Upscale to the resolution you actually need for printing or display.
- Restore faces last, when the model has the cleanest and largest face it is going to get.
Colourisation sits outside that list deliberately. It belongs after everything else, and it is optional.
Stage 1: scan it properly, or nothing else helps
No model can recover detail the scan never captured. This table is the part most restoration guides skip, and it is the part that decides your final quality.
| Original | Scan at | Why |
|---|---|---|
| Print, A4 or larger | 600 dpi | A sensible floor for enlarging a large print |
| Print, small (postcard, wallet) | 1200 dpi | You are enlarging a small original, so it needs more pixels per inch |
| 35mm negative or slide | 2400–4800 dpi | Film holds far more detail than any print made from it |
| Medium format film | 1200–2400 dpi | Larger frame, so lower dpi reaches the same detail |
| A phone photo of a photo | Not recommended | You inherit glare, perspective distortion and sensor noise |
Four scan settings that decide whether stage 4 and 5 will work
- 48-bit colour TIFF as the master, not JPEG. JPEG artefacts are exactly what a denoiser then has to guess at.
- Turn automatic sharpening off. Scanner sharpening bakes halos into the file that upscaling will amplify.
- Turn automatic dust removal off unless you are scanning suitable colour film. Digital ICE relies on infrared and does not work with traditional silver-halide black-and-white film or Kodachrome.
- Keep the original untouched and store the master separately — every restored version should be a derivative you can throw away.
Stages 4–6: denoise, upscale and restore faces locally
This is the part Formatif AI covers, and it covers it in one place rather than three. The order below is deliberate: denoising before upscaling prevents the upscaler from faithfully enlarging grain and dust along with detail, and face restoration goes last because the model then works on the largest, cleanest face available.
- Denoise the master scan. Add the Denoiser to the workflow and load the TIFF you repaired in the previous stage. Keep the strength moderate — over-denoising removes skin texture and fabric weave along with the grain, and no later stage can put it back.
- Upscale to the size you need. Use the AI Upscaler at 2× or 4× depending on the target. For print, work out the final inches at 300 dpi first and upscale to that, rather than guessing at 4× because it is the bigger number.
- Restore faces. Apply the Face Restorer to the upscaled image. Use it at a restrained setting on faces you recognise — see the FAQ below for why this is the stage to be most sceptical about.
- Run it as a chain. Because these are steps in one workflow rather than three applications, an entire folder of scans processes unattended with per-file status, and a single damaged file does not stop the batch.
Requirement note: Formatif AI needs a DirectX 12 compatible GPU — NVIDIA GTX 900 series or newer, or AMD RX 400 series or newer — with 6 GB+ VRAM recommended. Batches of 1200 dpi TIFFs also want 16 GB of RAM to stay comfortable.
Stages 2–3: damage, dust and tone — where Formatif AI is not the answer
Two stages on this list have no local AI equivalent in Formatif AI, and it would be easy to blur that. Here it is plainly.
Scratches, tears and dust
These need healing and inpainting: Photoshop's Healing Brush, Clone Stamp and Content-Aware Fill, or GIMP with the Resynthesizer plugin, which is free and fully local. Formatif AI's Object Remover will take out an isolated spot or blemish, but a scratch running the length of a photograph is a different kind of repair.
One caution: aggressive scratch removal over-smooths faces and flattens texture. Repair the minimum and leave character alone.
Tone, colour casts and colourisation
Correcting a faded print is straightforward in any editor — levels, curves and white balance. Colourising black-and-white is not: Formatif AI has no colourisation model. For that you need Photoshop's Neural Filter Colorize, or DDColor / DeOldify run locally through ComfyUI.
And colourisation guesses. It produces a plausible past, not a historical record — worth knowing before you show the family.
Tool stacks compared
Since no single application covers all six stages, what you actually choose is a stack. These are the combinations people run in 2026.
| Stack | Covers | Cost |
|---|---|---|
| Topaz Photo + ON1 Photo RAW | Everything except colourisation | Hundreds, one-time plus upgrades |
| Photoshop + Topaz Photo | Every stage, including colourise | Adobe subscription |
| GIMP + Upscayl + ComfyUI | Every stage, with manual work | Free, but technical |
| Formatif AI + a healing editor | All three AI stages locally, editor for damage | $12.95 plus whatever editor you own |
| Cloud restore services | One-click all-in-one | Credits or subscription — and an upload |
If you already own Photoshop, the cheapest complete route is Photoshop for damage and colourise plus a local AI tool for the three cleanup stages. If you own neither and want to spend once, Formatif AI plus a free GIMP install gets you the same coverage for $12.95— with the trade-off that GIMP's healing tools are manual labour rather than a brush you sweep across a scratch.
A note on the AI quality question, since it is the obvious one: we do not claim Formatif AI beats Topaz or Photoshop at restoration. For a fuller comparison of that trade-off, see Formatif AI vs Topaz.
Cloud restore sites vs doing it locally
The one-click services are genuinely convenient and, on heavily damaged images, sometimes better than a manual pipeline. The trade is not subtle: MyHeritage, Remini, VanceAI and their equivalents process your photograph on their servers, which means the image is transferred and handled under their terms. For a snapshot of a beach holiday that may be fine.
Family archives are a different category of file, though. They are frequently the only surviving copy of a person, they often depict living relatives, and they occasionally sit alongside documents you would not email to a stranger. Local processing removes the question rather than managing it — there is no retention window to read and no policy to trust, because the file never moves.
The reasonable conclusion is not that cloud restoration is bad. It is that a local pipeline is available now, it costs less than a year of credits, and family photographs are exactly the material where “just upload it” deserves a second thought. More on the offline vs cloud breakdown.
Restoring old family photos: FAQ
How do I restore old family photos without uploading them?
Work in stages on your own machine. Scan the original at the correct resolution, denoise the scan, upscale it, and run face restoration — all locally. Scratches and colourisation are separate stages that need different tools. The reason this matters for family photos specifically is that cloud restoration services such as MyHeritage, Remini and VanceAI all require the image to be uploaded to their servers.
What dpi should I scan old photos at?
Prints: 600 dpi minimum, and 1200 dpi for small prints, because you are enlarging them and need the extra pixels. 35mm negatives and slides: 2400 to 4800 dpi, since film holds far more detail than a print made from it. Medium format: 1200 to 2400 dpi. Always scan as 48-bit colour TIFF with automatic sharpening and automatic dust removal turned off, and never work on the only copy.
Can Formatif AI repair a torn or scratched photo?
Not as a dedicated restoration tool, and it is worth being straight about that. Its Object Remover can take out isolated blemishes, spots and small unwanted objects. But creases, tears and heavy scratch patterns are the job of healing and inpainting tools — Photoshop's Healing Brush, Clone Stamp and Content-Aware Fill, or GIMP with the Resynthesizer plugin. Do that stage first, then run the AI cleanup on the repaired image.
Can Formatif AI colorize a black-and-white photograph?
No. Formatif AI has no colourisation model, and anyone telling you a denoise-and-upscale tool will colourise is describing a different product. For colourisation you need a dedicated model — Photoshop's Neural Filter Colorize, or DDColor or DeOldify run locally through ComfyUI. A practical tip: colourise only after the image is denoised and correctly exposed, because these models work from luminance and respond badly to faded, noisy input.
Is AI face restoration accurate?
It is plausible, which is not the same thing. Face restoration models reconstruct detail that they think should be there rather than recovering pixels that survived. On a 1970s snapshot, that can produce a strikingly sharp result that no longer looks quite like the person. Our recommendation is to use it at a low strength on faces you actually recognise, keep the unrestored version, and be sceptical of any tool that claims exact recovery.
Do cloud photo restoration services upload my photos?
Yes — it is inherent to how they work. MyHeritage, Remini, VanceAI and comparable online restoration services process your image on their servers, which means the photograph is transferred, stored temporarily and handled under their terms rather than yours. For most family snapshots that may be an acceptable trade. For photographs of living people, documents, or anything you would not email to a stranger, local processing removes the question entirely.
What hardware do I need for local AI photo restoration?
Formatif AI needs a DirectX 12 compatible GPU — NVIDIA GTX 900 series or newer, or AMD RX 400 series or newer — with 6 GB+ VRAM recommended. Integration graphics will not run the models. Scanners and large TIFF batches also reward memory, so 16 GB of RAM is a sensible floor once you are working with 600 to 1200 dpi scans.
Should I keep the original scan?
Always. Save a master TIFF or PSD with the untouched scan, and treat every restored version as a derivative. Two reasons: restoration decisions are subjective and you will want to redo them with a better model in a year, and AI stages such as denoising and face restoration can over-smooth skin, fabric texture and film grain in ways you only notice later.
Keep reading
Run the AI stages on a whole folder of scans.
Formatif AI handles denoising, upscaling and face restoration locally, chained in one pass over a folder, for $12.95 one-time with a 3-day full-featured trial. Bring a scan you already own and see what the three stages do to it.
Start with the scan quality. It is free to get right and it decides more of the outcome than any model you choose afterwards.
Written by
Founder of Formatif, a local-first offline media processing suite for Windows and Linux. Hadi writes about video, image, and audio formats, compression, and privacy-first desktop workflows.