My mother kept a box of photographs that spent forty years in a damp attic. Water stains, a diagonal tear across a wedding portrait, and a layer of dust that had practically fused to the emulsion. That box became my test lab for this article. I ran the same damaged prints through Photoshop’s AI restoration tools and through old-fashioned manual retouching, side by side, and tracked what actually held up under a full-resolution zoom. This is a hands-on guide to restoring old photos in Photoshop, built from that testing, not from a press release.
Which Method Actually Restores an Old Photo Better, AI or Manual Photoshop Work?
Short answer: neither wins outright. AI restoration in Photoshop, mainly the Photo Restoration Neural Filter and Generative Fill, handles roughly seventy percent of common damage in under a minute. The remaining thirty percent, the parts that actually determine whether a restoration looks convincing, still need a human hand on the Clone Stamp and Healing Brush. I call this split the 70/30 Restoration Rule, and I’ll come back to it later because it’s the single most useful thing I learned during testing.
Here’s the nuance that most tutorials skip. AI restoration is fantastic at removing texture damage: scratches, dust, grain, and faded color. It’s much weaker at reconstructing missing structure, like a torn-off corner that used to contain someone’s hand or the edge of a doorway. That distinction between texture damage and structural loss is the entire logic behind how you should approach any old photo in Photoshop.
What “Restoring a Photo” Actually Means
People use “restoration” loosely, so let me define it properly before we touch a single tool. Photo restoration in Photoshop covers four separate operations, and mixing them up is why a lot of restorations look off. Enhancement sharpens and color-corrects an already-intact image. Repair removes localized damage like scratches or spots. Reconstruction rebuilds missing content, torn corners, or faded-out sections. Colorization adds plausible color to black-and-white originals.
To make damage assessment consistent across my tests, I built a simple scale I’m calling the Damage Class System, DC-1 through DC-4. DC-1 is light fading and minor grain, the kind you fix with adjustment layers alone. DC-2 adds scratches, dust, and small tears under about five percent of the frame. DC-3 covers major creases, mold spotting, and missing corners up to twenty percent of the image. DC-4 is severe: half the photo is physically gone, the emulsion is peeling, or there is water damage that has erased entire faces. Your restoration strategy should change completely depending on which class you’re dealing with, and I’ll reference this scale throughout.
The Two Restoration Paths in Photoshop, and a Third Hybrid One
As of August 2026, Photoshop gives you three distinct routes for old photos, and each one is genuinely suited to a different damage class.
Path One: the Photo Restoration Neural Filter
Found under Filter, Neural Filters, Photo Restoration. It’s built specifically for scanned prints and runs entirely automatically. Three sliders do the work: Photo Enhancement for color and contrast, Enhance Face for sharpening facial detail, and Scratch Reduction for removing physical damage. It’s genuinely good at DC-1 and DC-2 damage. It starts to struggle once you hit DC-3.
Path Two: Manual Retouching
Clone Stamp, Healing Brush, Spot Healing Brush, and Content-Aware Fill, the tools that have restored photos since Photoshop 7. Slower, but it gives you pixel-level control the AI filter simply doesn’t offer. For DC-3 and DC-4 damage, this is still where the real work happens.
Path Three: Generative Fill
Generative Fill uses Adobe’s Firefly model to invent new pixels based on a text prompt and the surrounding image context. It’s the only tool in Photoshop that can plausibly reconstruct a missing corner or a torn-away section rather than just cleaning up what’s already there. It’s also the tool most likely to hallucinate a detail that never existed, which I’ll get into shortly.
My Testing Setup: How I Actually Ran This
I scanned three prints from my mother’s box at 600 DPI, saved them as TIFF, and picked one DC-2 print (light scratches, faded color), one DC-3 print (a torn corner and a crease across a face), and one DC-4 print (roughly a third of the image missing due to water damage). Scan quality matters more than people assume. A tilted or low-resolution scan gives every AI model worse source data to work from, and Photoshop’s restoration tools are no exception. If you’re restoring old photos in Photoshop, spend the extra five minutes on the scan before you touch a filter.
Step-by-Step: the AI Neural Filter Restoration Workflow
Here’s the exact sequence I used, and it’s the one I’d recommend to anyone starting out.
Duplicate your background layer first. Neural Filters work destructively unless you’re on a copy, and you’ll want the untouched original to compare against.
Open Filter, Neural Filters, and enable Photo Restoration.
Start with Photo Enhancement at around 40, not the default maximum. Pushing it too far flattens contrast and gives skin an odd, waxy look.
Enable Enhance Face only if there’s a recognizable face in the frame. On group shots, it can overcorrect and make everyone look slightly too smooth.
Turn on Scratch Reduction last, and keep it conservative. This is where the filter does the most damage to legitimate detail if you push it past roughly 50 to 60 percent strength.
Export back to the main Photoshop workspace and inspect at 100 percent zoom before doing anything else.
On my DC-2 print, this alone fixed about eighty percent of the visible damage in under two minutes. On the DC-3 print with the torn corner, the filter cleaned up the surface scratches nicely but left the actual tear untouched, because Scratch Reduction repairs texture, not missing structure. That’s the exact limit the 70/30 Restoration Rule is describing.
Step-by-Step: the Manual Restoration Workflow
For anything DC-3 or worse, this is where I actually spent most of my time.
Use the Spot Healing Brush first for small dust and scratches. It samples automatically and is fast for isolated spots.
Switch to the regular Healing Brush for larger scratches near edges or faces, where you need to manually pick a clean source point.
Use the Clone Stamp for anything with a repeating pattern nearby, like fabric texture or a plain background, since it copies pixels exactly rather than blending them.
For the torn corner on my DC-3 print, I used Content-Aware Fill on a rough selection first, then went back with the Clone Stamp to correct the parts it got wrong.
Finish with Dodge and Burn on a separate layer to rebuild lost contrast and depth, especially around faces where the AI filter tends to flatten things.
This workflow took roughly forty minutes on the DC-3 print, against about ninety seconds for the AI pass. The extra time bought a noticeably cleaner result around the tear, particularly where the crease crossed a face. AI restoration in Photoshop just isn’t reliable yet at reconstructing a human face across a physical crease.
Where Generative Fill Wins, and Where It Hallucinates
Generative Fill earned its keep on my DC-4 print, the one missing roughly a third of the frame due to water damage. I selected the damaged area with the Lasso tool, left the prompt field blank so it would infer content from context, and let it generate three options. Two of the three were genuinely usable and blended almost seamlessly into the surviving edges.
Here’s where I’d push back on the more enthusiastic tutorials floating around. Generative Fill doesn’t know what actually used to be in that missing section. It’s predicting a plausible fill based on pattern continuation and on backgrounds, clothing, or plain surfaces that works beautifully. On faces, hands, and identifiable objects, it’s guessing, and the guess can be wrong in ways that matter emotionally, not just technically. I’d define this as the Hallucination Risk Zone: any area where the missing content includes a specific, identifiable feature a real person would notice is wrong. Never trust Generative Fill unsupervised inside that zone. Always cross-check against any other surviving photo of the same person or scene before you commit to a generated fill on a face.
Colorizing Black-and-White Photos: Is It Worth It?
The Colorize Neural Filter pairs naturally with Photo Restoration, and I tested it on an older black-and-white print from the same box. Results were solid for skin tones and skies, less reliable for anything unusual, like a specific dress color or a painted wall you actually remember. Treat colorization as an interpretation, not a fact. If accuracy matters to you, keep the original black-and-white version alongside the colorized one rather than replacing it entirely.
The 70/30 Restoration Rule
Across every print I tested, the pattern held steady enough that I’m comfortable calling it a rule rather than an observation. AI tools in Photoshop, the Neural Filter and Generative Fill together, reliably handle about seventy percent of a damaged photo’s problems: fading, grain, dust, minor scratches, and background reconstruction. The remaining thirty percent, structural tears crossing faces, fine hair detail, and anything in the Hallucination Risk Zone, still needs manual retouching to look right. Budget your time accordingly. Don’t expect a one-click fix on anything past DC-2, and don’t waste hours manually cloning out grain that a single Neural Filter pass would have handled in seconds.
AI vs. Manual Photo Restoration in Photoshop: A Side-by-Side Comparison
FactorAI Neural Filter / Generative FillManual RetouchingBest damage classDC-1 and DC-2DC-3 and DC-4SpeedUnder 2 minutes typically30 to 90+ minutes depending on damageFace reconstruction across tearsUnreliable, high hallucination riskReliable with practiceFine texture (grain, dust, fading)ExcellentSlow but preciseMissing large sectionsGood on backgrounds, risky on subjectsTime-consuming but controllableLearning curveLowModerate to highCostIncluded in most Creative Cloud plans via generative creditsIncluded, time is the real cost
What I’d Predict for Photo Restoration in Photoshop by 2027
Based on how fast the Neural Filter and Firefly models have improved since their 2023 launch, I’d bet on three things happening within the next year. First, face-aware restoration will get meaningfully better at handling creases and tears across facial features specifically, since that’s the most requested and most complained-about limitation right now. Second, expect tighter integration between Generative Fill and reference photos, letting you feed in a second undamaged photo of the same person so the model has something real to match against instead of guessing. Third, I expect Adobe to add a confidence indicator on generated fills, flagging low-confidence regions so users know exactly where a result is closer to invention than restoration. None of that exists yet, but the direction of travel points there.
My Honest Take After Testing Both Methods
If someone hands me a lightly faded photo with a bit of dust, I’m reaching for the Neural Filter every time, and I won’t feel bad about it. It’s fast, it’s genuinely good, and manually cloning out grain by hand is not a good use of anyone’s Sunday afternoon. But the moment a photo has a real tear across someone’s face, or a section is missing where a specific person used to stand, I close the AI panel and open the Healing Brush. The stakes are different when you’re restoring the only photo of someone who’s gone. Guessing isn’t good enough there, and right now, AI in Photoshop still guesses more than most people realize.
Common Questions on Restoring Photos in Adobe Photoshop
Is Photoshop’s Photo Restoration Neural Filter free to use?
It’s included with any Creative Cloud plan that has Photoshop, at no extra cost beyond your subscription. It doesn’t consume generative credits, unlike Generative Fill.
Can Generative Fill restore a torn or ripped photo?
Yes, for backgrounds and plain surfaces it works well. For faces or identifiable objects inside the torn area, treat the result as a guess and verify it against other photos before trusting it.
What DPI should I scan old photos at before restoring them in Photoshop?
600 DPI is a solid minimum for standard prints. Go up to 1200 DPI for small wallet-sized photos you plan to enlarge. Save as TIFF, not JPEG, to avoid compression artifacts stacking on top of the original damage.
Is AI photo restoration in Photoshop better than manual retouching?
Neither is universally better. AI wins on speed and light to moderate damage. Manual retouching wins on accuracy for severe damage, especially anything involving faces or missing structural content.
Can I restore old photos without Photoshop?
Yes, dedicated restoration apps exist, and some do a competent job on light damage. None I tested matched the combination of the Neural Filter plus manual layer-based retouching for anything past moderate damage, mainly because you lose the non-destructive editing control Photoshop gives you.
Browse WE AND THE COLOR’s AI, Technology, and Photography sections for more.
Subscribe to our newsletter!
The post How to Restore Old Photos in Photoshop: AI Neural Filters vs. Manual Retouching, Tested appeared first on WE AND THE COLOR.