Runs on WebGPU · no uploads · free

Zoom. Enhance.

16× AI upscaling that runs entirely on your GPU, inside a browser tab. No uploads. No account. No subscription. No install. Just drop an image.

16×
max upscale
3.1
TFLOP/s on a laptop GPU
0 B
uploaded. Ever.

apollo11-aldrin.jpg · 144 px → 576 px

Helmet enhanced 4× by lucent
Original lucent 4×
Pre-rendered with lucent. Checking for WebGPU…

That's a 72 × 72 pixel visor.

Upscaled 16× in two AI passes, by the same code that runs in your browser. The reflection holds Neil Armstrong and the Lunar Module. (The model reconstructs plausible detail. Great for photos, not for evidence. Sorry, CSI.)

Everything a desktop upscaler does

Minus the download.
Minus the upload.

16×

Upscale 1× to 16×

Presets or an exact pixel size. Past 4× lucent chains two AI passes, with a gentler second pass so detail doesn't turn to plastic.

Private by physics

There's no server to upload to. Images are processed by your GPU and never leave the tab.

network requests with your photo0

Live, tile by tile

Every slider re-runs the model, starting with the tiles in the centre of your view. The image never blanks.

Real denoise control

The denoise slider blends two networks' weights, not just their outputs.

Batch it

Drop fifty images, hit Export all, and they go straight into a folder.

Photos and art

A model for real-world photos, another for anime, illustration and line work. Plus sharpen, film grain and a "preserve original" blend for when the AI gets too creative.

How it works

A neural network, hand-wired
into your graphics card.

  1. 01

    Tile

    Your image is cut into 512 px tiles, each padded with 24 px of real neighbours so the seams vanish.

  2. 02

    34 layers

    Real-ESRGAN's compact network runs as 34 hand-written WGSL compute kernels in fp16, 3 TFLOP/s.

  3. 03

    Reshape

    Pixel shuffle to 4×, then Lanczos to your exact size, blended and sharpened.

  4. 04

    Paint

    Tiles stream back while the GPU starts the next one, nearest-to-your-view first.

Benchmarks, measured, not vibes

Fast because every
kernel was earned.

first kernel0.59 TFLOP/s
register blocking1.21
weights in shared memory1.79
fp163.13 TFLOP/s
0.5 s0.4 MP photo → 4×
6.9 s6.4 MP → 25.6 MP
5.3×faster than our first kernel
0/255error vs. the reference (fp32)

NVIDIA Turing laptop GPU, Chromium. Method and raw numbers live in the repository's docs/PERFORMANCE.md.

The honest table

Pick your trade-off.

lucentCloud upscalersDesktop apps
Your photos leave your deviceNeverUploadedNo
InstallNone, it's a tabNoneHundreds of MB
AccountNoUsuallyUsually, for the licence
Price$0Credits / subscriptionPaid licence, typically
Works on a phoneYes, with WebGPUYesNo
Face recovery & videoNot yetVariesOften

Pricing

Your GPU already paid for this.

Hobby

$0/forever

  • 16× upscaling
  • Both AI models
  • Unlimited images
  • PNG, JPEG, WebP
Start free

Enterprise

$0/forever

  • Everything in Pro
  • SSO: you're already signed in to your laptop
  • Data residency: your desk
  • SLA: as good as your fans
Contact your GPU

FAQ

Questions, answered.

Is it really free? What's the catch?

No catch. All the compute happens on your hardware, so there is no server bill to pass on. The app downloads about 6 MB of model weights once (cached afterwards) and that's the only traffic.

Where do my images go?

Nowhere. They're decoded, enhanced and encoded inside the browser tab. There's no backend to send them to. Open your browser's network panel and watch.

Which browsers work?

Anything with WebGPU: Chrome, Edge and Brave on desktop, Safari 26+, and Chrome on Android. Half precision (fp16) is used where the GPU supports it, for about twice the speed.

How big can images get?

Up to 16× the input size and about 120 megapixels of output, the most a browser canvas holds reliably. Above 4×, lucent runs two AI passes.

What model is it?

Real-ESRGAN's compact models (BSD-3-Clause, by Xintao Wang et al.), converted to fp16 and run on WebGPU kernels written for lucent. No ONNX Runtime, no WASM.

Is the added detail "real"?

It's reconstructed: the network predicts plausible texture and edges from what's there. That's what makes it look great, and it's also why you shouldn't use it as forensic evidence.

Make it lucent.

Open the app

No sign-up. It's already running.