feat: five of the seven Neural Filters run a network

[IAmJSD]

The README said two of them did. Now five do, and the two that do not say why.

What is new

Trained here, shipped inside the binary (tools/train/, and each script's defaults reproduce the weights that ship):

	params	what it does
dejpeg.onnx	66k / 256 KB	+1.0 to +1.2 dB over the JPEG at every quality from 10 to 60, on held-out images. Trained on Kodak compressed across that whole range, with the patches cut at unaligned offsets so it has to find the 8-pixel grid rather than assume where it is — a filter runs on a selection that starts anywhere.
colorize.onnx	466k / 1.8 MB	Predicts a distribution over 249 binned chromas rather than a chroma. Asked for one number, a network that cannot tell whether the car is red or blue answers brown, which is how colourisers end up sepia; class rebalancing and an annealed mean make it pick a side. Trained on 20,000 CC BY 2.0 photographs from Open Images. 1.1× the colourfulness of the real thing on held-out photographs, at 0.43 chroma correlation.

Downloaded on demand, hash checked, both running in tract unmodified: MiDaS v2.1 small (66 MB, MIT) for Depth Blur, UltraFace RFB-320 (1.2 MB, MIT) for Skin Smoothing. Depth Blur is now a real defocus — subject sharp and street gone, or the reverse — instead of a guess from local sharpness. Skin Smoothing smooths skin that is on a face, and scales its radius to the face, instead of anything skin-coloured anywhere.

Super Zoom is retrained, because finding out why the deblocker would not train explained why Super Zoom underperformed as well. PyTorch's default convolution initialisation shrinks activations by about a third a layer, so a residual network with a zero-initialised tail can train for thousands of steps and sit at exactly the identity — scoring precisely as well as doing nothing, which is a very quiet way to fail. Kaiming initialisation and normalisation fix it: +0.40 → +0.58 dB at the same 39k parameters.

Colour Transfer keeps no model and is not missing one. Moving one image's colour distribution onto another is arithmetic; Photoshop's network there is doing the segmentation around it, which is Object Selection's problem.

Supporting changes

• schist-neural gains an Input enum separating models that change an image (tiles at full resolution) from models that answer a question about one (one resampled frame, letterboxed for the detector), per-channel input normalisation, and depth/faces/colour modules that decode each network's answer.
• Chroma is resampled up guided by the luminance the model saw, so colour stops at the edges of the subject instead of bleeding over them.
• Each filter memoises its network's answer against a fingerprint of the pixels — where the faces are does not change while somebody drags Strength.
• Manage Models lists the new entries with sizes and licences, and a download now reads Downloading… 12.4 of 63.7 MB rather than sitting on an ellipsis for a minute.

Checked

Full workspace suite green, clippy and rustfmt clean. New tests cover the deblocker against a real JPEG fixture, the colouriser by correlation rather than error (a timid colouriser scores better on error and looks broken), and depth and faces where those models are installed — skipped, not failed, where they are not. The app was also driven headlessly: the 66 MB depth model removed and re-fetched through the dialog, and a Depth Blur preview reporting Using Depth (Depth Blur) · MiDaS v2.1 small, Intel ISL, MIT.

Known limits, stated in the README

The colouriser is small enough to be confidently wrong: it painted a red patch into a river in one test image. It knows sky, foliage, wood and skin, and hedges on anything whose colour is a choice. UltraFace has opinions about dogs.

🤖 Generated with Claude Code