Local AI · Runs in your browser

Four times the size.
A real step up.

SRNet is a compact super-resolution model — about 1.25 million parameters — trained on clean bicubic-downscaled pairs. It predicts a residual on top of a bicubic ×4 upscale, so it starts from plain bicubic and scores 30.28 dB against bicubic's 28.23 dB on DIV2K validation. Everything runs on your own hardware: no uploads, no queues, no rate limits. Output is typically a touch smoother than a true high-resolution original.

Upscale factor×4

PSNR vs bicubic+2.05 dB

Parameters~1.25M

RuntimeWebGPU / WASM

Privacy100% on-device

CharacterA touch soft

Source image

Drop an image here or click to browse · PNG, JPEG, WebP, BMP
You can also paste with Ctrl V

Transparent images are flattened onto white before upscaling.

Inference

Backend

Input limit

Tile size

384 px

Edge context (min 38 px)

38 px

Show tile map Reveal how the image is processed in blocks

Each tile is padded with at least 38 px of surrounding context — the receptive-field margin this model needs for tiled output to be pixel-identical to a whole-image run. Smaller tiles use less memory; larger tiles are faster.

Idle

Model facts

Architecture, training & limits

Architecture: 3×3 conv head → 16 residual blocks (64 channels) → conv to 48 channels + PixelShuffle ×4, added to a bicubic ×4 upscale of the input. The final conv is zero-initialised, so an untrained network outputs exactly bicubic. Every convolution runs at input resolution — roughly 2.5 MFLOPs per input pixel.
Training: DF2K (DIV2K + Flickr2K) high-resolution images; antialiased bicubic ×4 downscale quantised to 8-bit, with no noise or compression. 16,157 steps, batch 32, 64×64 low-res patches, AdamW with a 4e-4 peak and cosine decay, L1 loss with MSE for the final 8% of the schedule, EMA weights, fp16 — about 30 minutes on two T4 GPUs.
Validation: 100 DIV2K validation images, full images, bicubic ×4 degradation, Y-channel PSNR with a 4 px border shave, 8-bit outputs: bicubic 28.23 dB, SRNet 30.28 dB (+2.05 dB).
Metric note: Measured with an antialiased bicubic downscale rather than the official MATLAB imresize, so the two numbers compare directly with each other but only approximately with published tables. Published tables usually report RGB PSNR, which reads about 1.5 dB lower than Y-channel PSNR on the same images.
ONNX: srnet_x4.onnx takes float32 NCHW RGB in [0, 1] with dynamic batch, height and width, and returns values to clip to [0, 1] — the same file this studio loads.
License: MIT for the code and weights. Training-data terms are separate: DIV2K is distributed for academic research use and Flickr2K images keep their original Flickr licences — confirm those before commercial use.

Trained only on clean bicubic-downscaled inputs: noise, blur and JPEG artefacts are upscaled rather than removed, and the model does ×4 only.

Waiting for an image

Bring an image to life

Drop a photo, screenshot or render on the left. SRNet produces a ×4 version at +2.05 dB PSNR over bicubic on DIV2K validation, running on your own GPU — nothing is uploaded.

01Choose an imageSmall images are ideal; large ones are tiled automatically.

02Run the modelWebGPU when available, WASM everywhere else.

03Compare & exportDrag the seam, zoom to real pixels, save the result.

Four times the size.A real step up.

Bring an image to life

The model isn’t loaded yet

Four times the size.
A real step up.