AI Face Enhancer — Restore Old Photos in Your Browser
Fix blurry photos of old family members, restore mugshots, sharpen video screenshots. Runs locally with GFPGAN — no upload, no signup, no watermark.
Drag a photo here, or click to choose a file.
JPG, PNG, WebP — works on portraits, group shots, and scans.
About this tool
WRRK's face enhancer runs an ONNX-converted GFPGAN model directly in your browser using ONNX Runtime Web. The pipeline detects faces in the input, crops each one, runs the face-specific restoration model, and pastes the result back into the original at full resolution. It prefers the WebGPU backend for hardware acceleration and falls back to WebAssembly when WebGPU is unavailable. Tile-based inference keeps memory bounded so the tool handles group shots and high-resolution scans without crashing the tab.
Common use cases: old family photos where the faces have softened with scans or age; video screenshots grabbed from low-bitrate footage where features are mushy; mugshots and document scans where compression has destroyed detail; and profile pictures resized too aggressively for a small avatar slot. The browser-only architecture means you can run it on family archives or sensitive documents without worrying about uploads. The tradeoff is the one-time 80 MB model download. After that, every restoration is local, free, and unwatermarked.
Not just faces? For full-image super-resolution try our AI Image Upscaler which uses Real-ESRGAN for whole-photo enhancement.
How to enhance a face (5 steps)
- Upload a photo. Drag a JPG, PNG, or WebP into the drop zone, or click Choose image. The file stays on your device — nothing is uploaded.
- Pick the strength. Toggle 2× or 4× face restoration. Use 2× for already-decent photos that just need polish; use 4× for very blurry or heavily compressed faces.
- Click Enhance. On the first run, the GFPGAN model (about 80 MB) downloads and caches in your browser. Progress is shown in the bar.
- Compare before & after. Drag the split-view divider to compare the original and the restored version. Check eyes, skin texture, and hair edges.
- Download the result. Save as PNG (lossless) for archival or printing, or JPG (smaller) for sharing. No watermark, no quota.
When to use it
- Restoring old family portraits for prints or albums
- Sharpening blurry profile pictures before posting
- Cleaning up faces in video screenshots and stills
- Reviving faded scans of vintage photographs
- Improving group shots where individual faces are soft
- Polishing personal-record photos (not for govt submission)
Frequently asked questions
+−What's GFPGAN?
GFPGAN (Generative Facial Prior GAN) is an open-source AI model from Tencent ARC Lab that's specifically trained to restore real-world face images. Unlike a general upscaler, it understands facial structure — eyes, nose, mouth, hair, skin — and rebuilds plausible high-frequency detail in those regions. We run an ONNX-converted version directly in your browser via ONNX Runtime Web with WebGPU acceleration where available.
+−Does it work on multiple faces?
Yes. The pipeline detects each face in the image, enhances it independently, and pastes the result back onto the original at full resolution. So group photos, family pictures, and event shots all benefit — every detected face gets the same restoration treatment.
+−How long does it take?
On a modern laptop with WebGPU, a single-face image takes about 3–8 seconds. A group photo with 5–10 faces might take 15–30 seconds because each face is processed sequentially. On WebAssembly fallback (no GPU), expect 3–5× longer. The model itself only downloads once (about 80 MB) and is cached after that.
+−Privacy — does my photo get uploaded?
No. The entire pipeline — face detection, enhancement, paste-back — runs in your browser using ONNX Runtime Web. Nothing is sent to any server. That makes the tool safe for personal photos, family archives, document scans, or anything else you wouldn't want sitting on someone else's machine.
+−Why does it sometimes change features?
GFPGAN is a generative model — it invents detail from a learned prior of what faces look like. On very low-resolution or heavily blurred inputs, it may guess wrong about subtle features (eye shape, jawline, age). The result is always a plausible face, but it's not a forensic reconstruction of the original. If a feature looks off, try a milder enhancement or a different crop.
+−Can I use this for ID photos?
For personal records — restoring an old passport scan for a family album, sharpening a faded driver's license photo for your own files — yes. For official government submission (visa, passport, Aadhaar, PAN, etc.) — no. Because GFPGAN modifies the actual face pixels with generative detail, the output is not legally the same image as the original. Always submit unedited, properly-photographed images for official documents.
+−Does it work on group photos?
Yes — group photos are one of the best use cases. Each detected face is enhanced individually and pasted back onto the original-resolution image, so the background, clothing, and composition stay exactly as they were. You get sharper faces without any change to the rest of the photo.
Related on WRRK Tools
Flip images horizontally or vertically
ToolBlur images or selected regions
ToolRender SVG to PNG at any size
CategoryBrowse the full image tools category
ArticleDiscover how to compress images without sacrificing quality. We'll walk you through the best methods, tools, and settings to shrink file sizes while keeping your photos sharp.
ArticleHEIC files from iPhones won't open on Windows by default. Here are the fastest ways to convert them to JPG format that works everywhere.
PresetStamp-size profile pictures