Research Lab · V3.10

Advanced adaptive fusion

One crop, one run. SwinIR leads in natural texture; Real-ESRGAN contributes more where its local structure is clearer. SWINIR now uses the separately validated 128 × 128 ONNX model, while Real-ESRGAN retains its original 64 × 64 tiles. If WebGPU fails during inference, the app automatically retries the complete model pass using WASM. Models run locally in your browser; no API key, subscription or per-image fee.

1 · Load RGGB image

SWINIR ×4 model file · select the ONNX file once; the app will remember it locally

SWINIR model not selected. Choose swinir_realsr_x4_128.onnx (about 125 MB). The app saves a browser-local copy for future visits when Safari allows it.

Choose an RGGB-debayered image. First test one model at a time; the status will expose the full browser error.

2 · Select crop area

Choose a square crop, then drag it over the terrain. Crop size is how much of the original photo you select. It does not change the model. Compare both SWINIR models using the SAME 128 × 128 crop. Original SWINIR uses 64-pixel internal tiles; new SWINIR uses 128-pixel internal tiles.

Source crop size

Pinch to zoom · drag inside the crop to reposition it · drag outside the crop to pan the image.

Choose an image to position the crop.

3 · Compare identical crop

Original RGGB crop

Waiting for image

Real-ESRGAN ×4

Not processed

New SWINIR · 128-tile ×4

Not processed

Original SWINIR · 64-tile ×4

Not processed

Selective regional recovery · experimental

Run original SWINIR first

Advanced adaptive fusion

Not combined

Selective regional recovery is an experimental second pass of original 64-tile SWINIR using a denser, offset tile grid. It compares local edge response and only blends regions where the second pass has a stronger response. The original SWINIR result is kept unchanged.

Research caution: the combined image is an experimental blend. Original pixels remain unchanged.