One measure, many pipelines.

Upscale images and video, structure-first, no invented detail.

Contrast and sharpness placed exactly where φ sees structure.

Removes sensor noise and grain while φ preserves real detail — a small model that keeps structure sharp.

Astrophotography: φ lifts the light-pollution and vignetting gradient while preserving stars and nebulosity.

Frame-by-frame upscaling, encoded to MP4 in your browser.

Higher zoom factors, run offline on your own hardware — 4.66 MB model, φ inside. See the ×8 previews in the gallery.

The in-browser model learns synthetic noise. This one is trained on real sensor noise — signal-dependent, spatially correlated, chroma-dominant. Same 0.71 MB model, φ inside: on real photographs it gains +3.4 dB over the in-browser one, up to +7.7 dB on the noisiest shots. See the pairs in the gallery. Runs offline on your own hardware.

Upscaling while you watch, on your own graphics card: ×2 in real time (360p to 720p) and ×4 in real time (to 1080p, 27 frames per second). On Big Buck Bunny, measured against the original 1080p film: 4 to 6 dB above the video upscaling built into consumer graphics cards, which itself scores below a plain player upscale.

Removes the blocks and mosquito noise of MPEG-2, MPEG-4 and MJPEG video, then upscales. Ahead of the best classic chain in all 16 cases tested (4 films × codecs and qualities, +0.15 to +0.42 dB), and 0.3 to 1.1 dB above a plain upscale, with less flicker.

Upscale ×4 and sharpen a soft source at the same time. Under strong blur, the best result we measured: 2.2 to 3.3 dB above a plain upscale on four films, ahead of IMDN, a reference published network. Real-ESRGAN, though trained for blur, came last every time — it repaints faces instead of restoring them.

Structure detection validated in our publications, across visible and infrared. Not yet in-browser — available on request.

Semantic-free regions of interest, driven by φ alone — published and functional. Available on request.

A lossless and bounded-error file format that checks itself: every file carries the checksum of the decoded data, so an error is always detected, never returned silently. Against JPEG XL at its strongest setting (effort 9): infrared video from fixed cameras 36 % smaller, depth maps 7.5 % smaller, X-ray radiographs 7 % smaller, on par on still infrared and camera photos, with compression 5 to 12 times faster and 4 to 5 times less memory. Against the space standard CCSDS-123: infrared images 14 % smaller. Raw side-scan sonar: 3 % smaller than JPEG XL losslessly, and up to 41 % smaller with a guaranteed relative error of a few tenths of a decibel. With a guaranteed maximum error below the sensor noise: half the size of JPEG-LS near-lossless on infrared video. Decodes on a microcontroller in 57 KB of working memory. JPEG XL stays ahead on CT, MRI and ultrasound. Measured on public datasets (LTIR, ARCO, TUM RGB-D, Starosolski medical, CLIC, raw.pixls.us, AVIRIS, JNCC side-scan sonar, ARIS imaging sonar), every file decoded and checked.

Structure-aware processing for microscopy, astronomy, remote sensing.

A structural score to rank and audit image pipelines.

Radar archives kept only as colour maps have lost the rainfall itself: each colour hides a whole range. We recover the rain inside each colour, never outside it. On a real archive and a whole year never seen: 28 % less error than the best colour-by-colour reading, more than 2.5× the gain of the classic restoration method, yearly rainfall total within 1.5 %. It matches a state-of-the-art network with 32× fewer parameters, 1.8× faster on an ordinary CPU; a high-accuracy version goes beyond it. Validated on German, US and French (Météo-France) data, winter and summer storms, JPEG-compressed archives.

Real-time upscaling for retro and homebrew games. A 0.47M-parameter pixel-art model beats the best classic upscaling shader by 2.08 dB (PSNR, 200 test tiles), and the free model — 25 KB — holds 60 fps on the CPU alone. Our ×4 photo model ships inside it too: on clean game frames it outscores models trained on games, +10.2 dB on Freedoom. Small models learn structure, not a domain.