AI Face Restorer

The AI Face Restorer is a free, browser-based tool that sharpens faces in old, blurry or low-resolution portrait photos using the Swin2SR neural network. The model runs entirely on your own device, so photos are never uploaded, and it restores detail in under 30 seconds where studio restoration costs $30–$200 per image.

How Swin2SR AI Face Restoration Works

This AI Face Restorer uses Swin2SR — a Swin Transformer V2-based image restoration model from Microsoft Research — running via Transformers.js in your browser. Unlike simple Canvas sharpening, Swin2SR understands image structure and reconstructs fine detail genuinely lost to age and scanning: facial features, hair strands, skin texture edges. The model is downloaded once (~50MB) and cached for offline use. After the AI pass, a portrait-specific post-processing pipeline applies gentle noise reduction and targeted sharpening calibrated for skin tones. Last updated: March 2026.

The restoration pipeline: your photo is resized to max 512px before the AI (keeping processing fast), fed through Swin2SR for super-resolution reconstruction, then post-processed with contrast and sharpening at the mode level you select (Gentle / Standard / Strong). If the AI model is unavailable the tool falls back to the full Canvas pipeline automatically — no broken experience.

Restore Old Family Portrait Photos

Family portrait photos from the 1940s–1980s typically suffer from fading (compressed tonal range), soft focus, and colour cast from yellowing paper. Swin2SR was trained on paired degraded-to-clean images, making it highly effective at recovering facial detail that Canvas filters cannot reconstruct. Professional portrait restoration services charge $30–$200 per photo. This tool is completely free and processes in under 30 seconds. Pair this tool with the AI Photo Colorizer to also add colour to black and white portraits — restore first, then colorize.

Portrait Enhancement Modes

• Gentle mode: Minimal post-processing on top of AI (contrast 8, light noise reduction, sharpAmount 0.15). Best for photos that are already reasonably clear — let the AI do the heavy lifting without over-processing.
• Standard mode: Balanced post-processing (contrast 15, moderate noise reduction, sharpAmount 0.3). Best for most old portrait photos — restores contrast, smooths grain, sharpens key features.
• Strong mode: Maximum post-processing on top of AI (contrast 25, strong noise reduction, sharpAmount 0.5). Best for heavily deteriorated photos — use with care on recent or already-clear photos.

What Resolution You Get Back — and How to Get More Detail

This is the part most browser restorers do not tell you. Swin2SR is a super-resolution model, and to keep it fast enough to run on a phone the tool caps the input at 512px on the longest edge before the AI pass. The model then reconstructs at up to 4× that, so a small photo comes back larger — but a 3000px scan is downsized first and comes back smaller than you uploaded. The result caption now prints both sizes so you can see exactly what happened rather than discovering it after download. The practical fix is simple: crop tightly to the face before uploading. A 480×480 crop of a head keeps every pixel of facial detail inside the 512px budget, so the AI spends its whole reconstruction on eyes, mouth and hair instead of on background wall. For a full-frame group photo, crop and restore each face separately, then reassemble. The Swin2SR architecture and its degradation training pairs are documented in the original paper on arXiv:2209.11345. Last updated: September 2026.

Face Restoration vs General Photo Enhancement

General photo enhancement boosts contrast, saturation, and sharpness uniformly. For landscapes and products that works well. For portraits, uniform sharpening creates "halo" artefacts around hair and clothing. The AI Face Restorer uses portrait-calibrated post-processing: lower contrast amounts that protect skin gradients, radius-limited sharpening that targets small facial detail (faces typically 200–600px wide) while leaving large background elements unchanged. The result is sharper eyes and lips without the "over-sharpened" ring artefact.