The model ledger

• ESPCN

  onnx/models super-resolution-10, opset 10

  SHA-256 85f36ff88cc504a24af5e0602148bc56a8aa09a58eca8c0da2756f3e8186035e

  Source: https://github.com/onnx/models — validated/vision/super_resolution/sub_pixel_cnn_2016

• Real-ESRGAN general x4v3

  realesr-general-x4v3, re-exported fp32

  SHA-256 ef6418714b70efbcec6eecbb0e4fccd1b9a34837f5bac5dcc3ffbb0ac2bea79e

  Source: https://github.com/xinntao/Real-ESRGAN, re-exported from the official checkpoint

• YuNet face detector

  opencv_zoo face_detection_yunet 2026may

  SHA-256 ebafce4e3c118d6554634be5c27ab333b4c047a9a8c3faf1d7cf93101c22f0f0

  Source: https://github.com/opencv/opencv_zoo/tree/main/models/face_detection_yunet

The runtime that executes them, 13,986,063 bytes of WebAssembly, is served from this domain. Weights are kept in this site’s own cache so a return visit fetches nothing; keep them here before you go offline, or remove them.

• None of them runs anywhere but this tab. If a browser cannot run WebAssembly, the studio and the queue fall back to Lanczos3, which needs no file at all — there is no server to fall back to.
• WebGPU is detected and reported, not used: the WebGPU build of the runtime is larger than this host will serve as one file, so every graph here runs on WebAssembly — on up to four threads where the browser isolates the page, one where it does not.
• Real-ESRGAN reconstructs. What it adds is plausible, not recovered — the studio's risk map exists to show where.
• YuNet finds faces and does nothing else. Its boxes steer the studio's face refinement and are never stored or sent.

• Photos upscalerUpscale one photo in the review studio, or a whole folder in the queue, on this device.
• Where the minutes come fromTile geometry read off both graphs, the per-pass figures, and the calibration that replaces them.
• PrivacyWhat leaves this device, what does not, and what is written to disk.
• Runs, warns, refusedThree bands, two thresholds, the halved iOS envelope and the arithmetic floor beneath all of it.