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 weights were in two places: face detection and recognition in
`models/face/`, segmentation in `core/dr-segment/models/`. Nothing was
wrong with either path, but between them there was nowhere to look to
answer "how much model does this application carry", and that number is
about to start growing.
So the crate-local copy moves up beside the other. `models/` now holds
`face/` and `segment/`, and a `du -sh` of one directory is the whole
answer.
No content changes: the .onnx and its vocabulary are byte-identical, and
`LICENCE.md` moves up a level to cover the tree rather than one crate.
The LFS pattern in `.gitattributes` is `*.onnx` and already matched both
locations, so only its comment needed the new path.
`include_bytes!` is relative to the source file and `build.rs` runs with
the crate root as its working directory, which is why the two paths climb
a different number of levels.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Local masking needs to know where an image's regions are. The watershed
spike (S15 arm A) found the boundaries but had no idea what any of them
enclosed; its coarse levels were geometric accidents. This adds the other
half and the thing that joins them.
`core/dr-segment` is where region reasoning now lives — the hierarchy moves
out of `dr-gpu`, which keeps only the pixel passes that are genuinely
shaders. The new crate is device-free and, without its default features,
model-free too: 20 of its tests need neither an adapter nor 11 MB of
weights.
Arm B runs YOLO26n-seg through `ort`. D13 framed inference as a choice
between `ort`'s C++ runtime and the pure-Rust dependency policy; that was a
false choice. `ort`'s `alternative-backend` feature unlinks the C entirely
and `ort-tract` supplies the API from tract, which is pure Rust. Measured
before committing to it: zero unsupported operators, 420 ms for 640x640,
and correct masks on bus.jpg. No NDK problem to solve, so D13's largest
tolerated exception is not needed.
Arm C is `prior.rs`, and it ships because the two arms fail in opposite
directions. Instance membership re-weights the merge saddles, so region
pairs the model believes share an object merge early and pairs straddling
its edge merge late. No boundary moves — only the order in which they
dissolve — which is how the result stays pixel-accurate at every level
while its coarse levels become named things.
Two things the spec assumed that turned out to be false, both recorded in
models/LICENCE.md: there is no usable ADE20K-trained YOLO, so the shipped
vocabulary is COCO's 80 subjects and *stuff* like sky and foliage must come
from arm A; and tract cannot parse a dynamic-shape export, so the graph's
input is fixed and tiling is the only route to more semantic resolution.
Weights are AGPL-3.0, which GPLv3 §13 permits and which makes the combined
work effectively AGPL. Deliberate, not accidental. They live in Git LFS,
and a build script fails with an instruction rather than embedding a
pointer file when the clone lacks them.
dtourolle
