QWE AI Academy You upload a scratched photo from 1982. Thirty seconds later, the face is sharp, the colors vivid, the tears gone. But your grandmother’s eyes look slightly different. Her smile shifted. You can’t prove it, but something’s wrong.
AI photo restoration doesn’t restore – it reconstructs. Sometimes it invents.
The Damage Test: What AI Can Actually Fix
Every tutorial lists features. We tested actual damage.
Eight AI restoration tools. Three damage scenarios: light scratches and fading (most photos), moderate tears and discoloration (common in scanned albums), severe damage with >50% facial data missing (rare but instructive). What worked, what failed, one limitation nobody mentions in the marketing copy.
Light damage – scratches, dust spots, faded colors. Every tool handled this. VanceAI, Remini, LetsEnhance, Hotpot – clean results in 10-30 seconds. For this tier? Pick whichever interface you like.
Moderate damage – torn edges, water stains, heavy discoloration. VanceAI ($5.94/month annual as of January 2026) and jpgHD handled tears better than faster tools. Remini excels at faces but leaves backgrounds noisy. MyHeritage ($7.83/month annual) required multiple passes: enhance, then colorize, then repair. Not one-click.
jpgHD’s catch: Old Photo Restoration mode takes 3-5 minutes (per official docs). Faces smaller than 256×256 pixels? Ignored. Group photo from across a room means some faces won’t be detected.
Severe damage – AI stops restoring, starts guessing. When more than half a face is missing or heavily degraded, the algorithm fills gaps with statistically plausible features. Community feedback on BestPhoto and Reddit documents the “Groucho Marx effect”: AI adding glasses, mustaches, entirely different expressions that never existed. We saw it. One test photo: sharper cheekbones, subtly different nose bridge. Plausible. Wrong.
For irreplaceable photos with severe damage, get the AI result AND manual restoration ($35+ per photo as of early 2026, per Photobooth and Rememorie). Compare them. AI shows what’s possible. Human shows what’s accurate.
Why does this happen? AI works by prediction, not recovery.
It doesn’t know which details are real – just which are statistically likely based on millions of training images. The more damage, the more it guesses. That’s the fundamental limitation of AI restoration: you’re trading completeness for accuracy. Sometimes you gain more than you lose. Sometimes you don’t.
Why ChatGPT Fails at Photo Restoration
You’ve seen the viral posts: someone asks ChatGPT to restore a family photo, and it comes back looking like a different person.
Not a bug. ChatGPT doesn’t restore images – it generates new ones based on your prompt. Reinterprets, not repairs. A Reddit user tried restoring the world’s oldest photograph (taken in 1826 by Joseph Niépce). ChatGPT added a church spire that never existed and changed the rooftop structure. Photography history teacher confirmed: “totally inaccurate” (per TechRadar June 2025).
This matters. Dedicated restoration tools like VanceAI or Hotpot analyze surrounding pixels – fill damaged areas based on local context. ChatGPT’s image generation operates at a conceptual level. It understands “old photo” and imagines what one should look like, ignoring the actual pixel data you gave it.
Want to preserve identity or historical accuracy? Don’t use ChatGPT. Use it for creative reinterpretations, not archival work.
The MyHeritage Pricing Trap
MyHeritage appears in every “best tools” list. Effective for portraits. Offers colorization. The comparison charts skip this:
The $7.83/month subscription (annual billing, as of January 2026) includes genealogy features, DNA matching, family tree tools, historical records access. Only need photo restoration? You’re paying for an entire platform you won’t use. The standalone Reimagine app: $49.99/year, focuses on photos – still more expensive than VanceAI ($71.28/year) or LetsEnhance ($108/year for 400 credits). Both offer more restoration features per dollar.
Use MyHeritage if you’re building a family tree. Just fixing photos? Better options exist.
Free Tier Reality Check
| Tool | Free Credits | Watermark? | What’s Actually Free |
|---|---|---|---|
| VanceAI | 5/month | Yes | Test quality, can’t use results |
| Remini | Limited | Yes | Few per day with ads |
| Hotpot | Unlimited | No | Fully usable (CC BY-NC license) |
| Photobooth.online | Unlimited | No | Fully free, no account needed |
| jpgHD | Limited | N/A | Old Photo Restoration not available |
Need watermark-free results without paying? Hotpot and Photobooth. Everything else requires a subscription for usable output (as of February 2026 per ScreenApp and BestPhoto comparisons).
Which Tool for Which Damage
Casual fixes (light scratches, fading): Hotpot or Photobooth.online. Both free, fast, watermark-free. Batch processing? Then pay.
Batch restoration (digitizing entire albums): VanceAI. Five free credits for testing, then $5.94/month (annual) for 100 credits. Handles scratches, tears, colorization in one pass. Saves hours.
Mobile-first workflow: Remini. Over 100 million downloads (per ScreenApp February 2026), best face enhancement on phones. Free tier: limited and watermarked. About $10/month for unlimited access.
Professional-grade control: Adobe Photoshop ($11.99/month Photography Plan as of February 2026). AI-powered Generative Fill and restoration filters, plus manual tools when AI fails. Topaz Photo AI used to be the upscaling king – switched to subscription-only at $58/month in October 2025 (per ScreenApp and BestPhoto February 2026). Most expensive option now.
Pattern nobody mentions: most tools use a single AI model for everything. One algorithm handles faces, scratches, colorization. Why some excel at portraits but fail on backgrounds, or vice versa. VanceAI and Adobe are exceptions – they let you choose different models or combine AI with manual edits.
When AI Can’t Help (And What to Do Instead)
Water damage. Mold. Photos where emulsion separated from paper backing. Extremely blurry originals shot on cheap cameras.
AI can’t invent detail that was never captured. Photo blurry because the subject moved during a long exposure in 1940? No algorithm will sharpen it – just guess at what details might have been.
For severely damaged or historically significant photos, professional manual restoration still costs around $35 per image minimum (as of early 2026, per Photobooth and Rememorie service sites). 6-7x AI tool cost per photo. You’re paying for human judgment: recognizing when a blur is damage versus period-accurate lens softness, preserving era-specific aesthetics, knowing when to stop.
Try this: scan at highest resolution your scanner supports (600 DPI minimum). Run it through an AI tool to see what’s possible. Result looks off – features changed, expressions shifted? Either try a different AI model or accept that manual restoration is the only path.
How do I know if AI invented details that weren’t there?
Compare restored version to original at 100% zoom. Look for asymmetry corrections (real faces are asymmetric; AI often makes them more symmetric), texture changes in clothing or backgrounds, expression shifts. Multiple photos of the same person from same era? Check if facial proportions stayed consistent.
If unsure, overlay the AI version on the original in Photoshop at 50% opacity. Features shifted position? That’s invention, not restoration.
Can I restore photos that are too damaged for AI?
Yes. Requires manual work. Photoshop’s clone stamp and healing brush tools let you rebuild missing areas by sampling from undamaged parts. Slow – expect 30+ minutes per photo – but you control what gets added.
Some professional restorers combine AI for initial cleanup, then manually correct areas where AI guessed wrong.
Do restored photos lose quality compared to the original scan?
Not if you use the right format. Always work from uncompressed originals (TIFF or PNG). Scanned photos are JPEGs? Convert to TIFF before restoration. AI tools often re-compress output as JPEG by default – degrades quality. Most tools let you choose output format. Pick PNG or TIFF for archival work, JPEG only if you need smaller files for web sharing.
Start with photos you care about least. Test tools. Compare results. Then move to the irreplaceable ones once you know what each tool does well – and where it fails.