denoise.md §14: why the tiles waste half of Best's work, where the any-size
networks run and why only there, why the limit is the card's memory, and
the measurement on _MG_8862 — 2.60 s in 1408 tiles, 1.37 s in two
4160 x 3248 tiles, the outputs within fp16's own spread. models/LICENCE.md
lists the three re-exports.
AI Denoise's Apply switch becomes Method: Bilinear, Fast, Medium, Best,
default Best, so an untouched raw writes nothing and develops through the
mixture. `apply` is still read and never written: 0 is Bilinear, 1 keeps
a network already chosen.
- Best is the mixture of a flat and an edge expert with a learned gate;
Medium and Fast are students distilled from it. 2.48 s, 0.79 s and
0.57 s for a 20 MP frame on TensorRT fp16.
- Each network carries its own tile border (256 for the mixture, 192 for
the students) through `dr_denoise::Shipped` and `TileNet::halo`.
- The file is hashed once at open and each network keys its own cached
result; Bilinear keeps the result in memory for the way back.
- Each has an .a16w16 sibling for the Hexagon: 0.00 dB on the 6D gate,
at most 0.11 dB with the noise scaled x0.5 to x4.
- APK BUNDLED 19 -> 23; the PKGBUILD installs all three.
The engine knew f32 and int8, and gave the Hexagon int8 for every role it
served. Measured on the tablet itself (inference.md §1.5), int8 lost
5% of the detector's faces at 40-80 px, moved the landmarks 1.5 px,
emptied the segmenter's scores and cost the denoiser 5-9 dB; fp16 the HTP
refuses outright. `Form` gains A16W8 and A16W16, and `Rung::form` now
names one per role: detectors and landmarks A16W8, the segmenter, scene
model, border filler and denoiser A16W16, XFeat int8. The embedder and
the eye classifiers stay on the CPU.
Each loader resolves its `<stem>.<form>.onnx` sibling; the segmenter and
XFeat, compiled into the binary, embed their quantised forms on Android
only and pick through `choose_embedded`. The probe, the compile step and
the cache fingerprint follow the form instead of assuming int8. Detectors
on the new form write `scrfd_*_a16+w600k_mbf`, and `model_ids` answers
for all three spellings.
On the tablet (ORT 1.29 + QNN 2.42), each shipped file against f32 on the
same inputs, and against the CPU's f32 time:
SCRFD 500m/2.5g/10g A16W8 100% of faces in every band 4.2/5.1/9.0 ms vs 17/56/198
landmarks A16W8 0.25 px in the 192 crop 0.5 ms vs 2.8
YOLO26n-seg A16W16 98.2% found, mask IoU 0.994 12.9 ms vs 90
scene model A16W16 98.9% of cells agree 15 ms vs 151
MI-GAN A16W16 41 dB from f32 in the fill 87 ms vs 488
XFeat int8 pano alignment 0.45 px (f32's own spread 0.41) 6.5 ms vs 58
denoiser A16W16 0.00 dB at every ISO 95 ms vs 1510 a tile
Face numbers are over public COCO val2017 photographs, not a library.
The APK carries the siblings (BUNDLED 15 -> 19; the old int8 detectors
removed), about 43 MB more. The Windows installer and its CI count skip
them; the Arch and Flatpak packages list their files and never had them.
The ladder example takes a role per model, which is how the per-role
forms above were seen landing on the NPU from the real probe.
The learned demosaic was an option under Detail, off by default. It is
now how a Bayer raw is developed: on by default at full strength on
every device — which hardware runs it is the inference engine's choice
— and first in the Adjust panel, since it decides what every control
below is applied to.
Strength (0-100, default 100) replaces Keep grain: grain = 100 -
strength, the same luminance-only blend, so moving it is one GPU pass
and never a re-run. 0.21.0's sidecars stored grain; it is still read,
as the inverse, and never written.
With it on for every photograph, the result is now kept on disk
(denoise.md §7.1, §12): the network's output as half floats, keyed on
a SHA-256 of the file's bytes and the model, oldest first past a 5 GB
budget, beside the inference engine's cache. A reopened photograph and
an export of one already developed read it back instead of running the
network again; a damaged entry is a miss.
The grain blend replaces the Amount of §7.2, and why its objection to a
blend does not hold for brightness alone; the Hexagon is out (int8 -6 to
-9 dB); §11 holds the data, the noise model taken from the library, the
model, the validation table, the blind estimate's reach and the speed.
A design draft for the learned stage outstanding.md §3 lists as missing:
a joint demosaic-and-denoise network that runs on the mosaic, in the
slot architecture.md §5.2 reserves for it, with its result kept as a
cache rather than a new file in the library. It covers the model and
tiling, synthetic training pairs from the library and a per-body noise
calibration, evaluation on real pairs, the Amount control, speed on the
tablet, X-Trans, and the decisions still open. Nothing in it is built;
figures marked "estimate" wait for the measurements that replace them.
dtourolle