Measured on a Radeon RX 7900 XT against Arch's onnxruntime-rocm 1.29
(docs/inference.md §1.3): MIGraphX fp16 runs the detectors at 2.4–3.4 ms
against 10–58 ms on the CPU provider, the inpainter at 8 ms against 514,
with a 15–135 s compile per graph the first time and under a second from
its cache after. A compiling rung on TensorRT's terms, wired the same way.
The ROCm execution provider is gone (removed in ONNX Runtime 1.23), so the
AMD ladder is MIGraphX then the CPU, with no non-compiling rung between.
MIGraphX is registered through the runtime's generic key/value entry
point rather than ort's builder: 1.29 reads the legacy options struct for
its precision flags only, and the compiled-program cache directory
(`migraphx_model_cache_dir`) only travels the generic way. The provider's
cache key omits the precision, so f32 and fp16 programs get their own
directories. The probe fingerprint now includes the provider libraries
beside the runtime and the ROCm version, since a distribution's CPU and
ROCm builds are the same file at the same path.
`status().failed` reports only the rungs above the selection, so an AMD
desktop's About line says why MIGraphX won rather than that the NVIDIA
providers are not in the build.
Two examples: `ep_probe` times each provider cold and from cache, and
`ladder` drives `init` as the app does to watch the first-run sequence.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A scene that makes a parent, nests two collections in it by drag and by
the menu, files frames into a child and opens the parent to see it count
both; stills of the tree and of the menu. The collections recording is
re-made now that the bitmap under the cursor is the photograph.
drive.py grows a multi-leg drag: a diagonal with much vertical in it is
taken by the grid's Flickable as a scroll before the DragArea can claim
it, so a drag to the sidebar goes sideways first.
`changed columns` fires on a change, and a first evaluation is not one:
a grid built after the window had settled at its size never said how
wide it was, so Rust placed month headings for the one column it was
told about at start-up — every month began a row, and was announced
wherever its first cell fell, mid-row included. Opening a collection
showed "October 2025" stranded over a row of August.
The bitmap under the cursor was a solid red rectangle. Slint's drag
overlay uploads the image as a texture, draws it and drops the texture in
one call; with the wgpu FemtoVG renderer the drop is immediate and the
draw is deferred to the flush, so the frame binds femtovg's placeholder —
which is red. An image with a cache key survives in the texture cache
until after the flush, and only a path gives one. So the composite goes
to the data directory's scratch as a PNG and comes back through
load_from_path; one file per drag, removed when the drag ends. A
workaround for Slint 1.17.1, written up as one beside the code.
The recording sampled a red brick wall and moved the sliders by three
units, which at GIF size is a click that does nothing. The scene now
drags the frame cold first and picks a white air conditioner, so the
correction is visible and the picker's being absolute - set from the
photograph, not from where the sliders were - is what the picture
shows. The text says so, and says a blown highlight is refused.
The probe's comment said a 192px render "averages a small neighbourhood
into each of its pixels". It does not: the composed shader fetches the
source at one position per output pixel - nearest for an unrotated
frame, four photosites blended otherwise - so the probe was a point
sample of a noisy sensor, and two painted-white air conditioners on the
same wall answered +37 and -50.
The tap is now narrowed to the patch of the canvas around the click, a
couple of percent of its width and square on screen, and rendered at
64x64 with interpolation forced on, which puts a sample on every sensor
pixel under it at any ordinary zoom. The samples are averaged, with the
void and clipped ones left out rather than allowed to pull the mean, and
fewer than half surviving is refused. compose_camera_probe takes the
patch; the merge's compose_camera_linear keeps its nearest sampling. The
readback shrinks from six megabytes to sixty-four kilobytes.
A frame of alternating warm and cool columns, averaging neutral, moves
the controls by at most two units; a point sample swung them to sixty.
The scene clicked 40px to the right of "pick", on "reset", so the
recording showed a neutral group being reset and a click on the wall
that panned. Re-recorded with the picker fixed: the word lights, the
sample moves temperature and tint, and Before shows what it corrected.
Sampling the overcast sky on a Canon 6D frame set tint to -100 and
temperature to -15 for a patch the canvas showed as pure white. A clipped
photosite is sensor white, not a colour: every channel stopped counting,
so what the tap hands back is the as-shot multipliers themselves, which
are strongly magenta, and the solver dutifully drove green to its stop.
The display shader already fades such a pixel to a neutral of the same
brightness before any operation runs, so the picker was balancing against
something the photographer could not see.
The probe now refuses a sample with any channel at or above the onset the
shader fades from, the way the solver already refuses black. The threshold
is one constant, CLIP_ONSET, formatted into the shader and read by the
probe, so the two cannot drift apart.
Pressing "pick" and clicking a near-neutral wall on a Canon 6D frame set
tint to -77 and turned the whole photograph green. The white balance
operation runs first in the chain, on camera RGB, before the body's base
curve and colour matrix; the probe was read off a display render after
all three, and the solve treated that sRGB triple as if the gains
multiplied it directly. On a JPEG the two spaces coincide, which is why
the existing tests passed while the picker was broken on every raw file.
The probe now reads the camera-space tap a merge stitches from, composed
under the edit's own framing so a fraction of the canvas is a fraction of
the probe, and puts the as-shot balance on itself - exactly the value the
operation's gains are about to multiply. No operations run in the tap, so
nothing has to be stripped and restored, and the display target is left
alone, so a sample that found nothing usable no longer needs a redraw.
A raw-frame test with the 6D's matrix and a typical as-shot balance
samples a warm grey and asserts the rendered pixel comes back neutral; it
fails on the previous probe.
Every face the user has ruled on entered the pass as an anchor, and the
scan is exhaustive by design (`dr_face::neighbours`), so a person with
750 confirmed faces cost 750 comparisons against every other face in
the library — and the cost of a library grew with how well it was
named. Most of those comparisons said nothing new: thirty frames from
one afternoon are one point of view, not thirty, and a face that
matches one of them matches the rest.
Each person now enters through at most 100 of their anchored faces
(`dr_face::references`). Eligible are those whose raw embedding is at
least 15 long — one above the gallery floor, since a reference speaks
for someone rather than merely being admitted — with an unmeasured
length admitted as it is everywhere else. From those, the set spanning
the greatest volume is chosen greedily: the longest vector first, then
at each step the face with the largest component orthogonal to the
chosen so far. That is pivoted Gram–Schmidt, and the product of the
residuals it picks is the Gram determinant, so the greedy step is the
exact greedy on the objective. A near-duplicate of a chosen face has
no residual and is passed over; the one profile shot among two hundred
frontal frames is taken early; faces inside the span of the chosen add
no volume and are not taken to fill the cap.
The faces not chosen keep their confirmations and are not touched by
the pass — they stay in the anchor map, so it never releases them —
they are simply not compared. A person none of whose faces is long
enough is still stood for, by their longest, rather than losing their
anchor and having their next face filed as a stranger. Under the cap
nothing changes: every eligible face stands, and the short ones stay
in as the probes they were.
At the reference library's 3,851 confirmations the scan shrinks by
about a fifth; at 15,000 it is a fifth of what it was.
The markers the previous commit stops writing are already in the
catalogs — 2 on the desktop, 429 on the tablet — and in the shards
both have exchanged. Renaming them to the faces' own id with a fresh
time is what makes the export send each image again, under an entry
newer than the empty one `held_model` would otherwise pick. Where the
old write had inserted its marker beside the right one, the wrong one
goes and the right one is refreshed for the same reason: its entry in
the shards is older than the empty one.
Images V14 left with faces and no marker are not touched. That state is
the quality pass's cue, and the fixed write marks them correctly when
it reaches them.
Checked against copies of both real catalogs: the desktop renames 2,
the tablet deletes 429, both in under 200 ms.
`record_updates` — the write behind the quality, eye and crop passes —
re-marked the image as indexed under the pipeline the pass ran as, and
left the faces it had updated under the id of the detector that found
them. On a desktop set to Thorough that put `scrfd_10g+w600k_mbf` over
faces spelled `w600k_mbf`; on the tablet, `scrfd_10g_i8+w600k_mbf` over
faces it had adopted from the desktop's thorough pass.
Every reader takes the marker and the faces to agree. `marker_under`
reads the marker as the detector having examined the image, so the
upgrade repair never revisits it. The shard store keys each face by
its pipeline id, so `export_to_shards` selects an image's faces by the
marker's id, finds none, and sends an entry that says the thorough
detector looked and found nothing — over photographs with named faces
on them. The desktop's shard index holds 54 such entries beside real
faces; the tablet's eye pass over the faces it had adopted made 430
more, and both devices have exchanged them. `held_model` takes the
newest entry for an image, which is the empty one. Nothing has been
lost yet only because the two spellings of the thorough detector rank
equal and neither side adopts the other's; a third device, or either
one after a reinstall, would adopt "nothing here" for 484 images. And
the desktop's eye pass is 4,739 images from doing the same to every
face from before V14 — which are the ones that only exist on the
desktop, and would then never reach anywhere.
The marker now takes the id the faces carry; the pass's own id is used
only when it dropped the last of them and there is no detector left to
name. A stale marker under another spelling of the same embedder is
removed in the same transaction, so one embedder has one marker.
The tablet showed a fraction of each person: 681 of the desktop's 3,851
confirmations, and none of Ian's 746, Catherine's 626 or my own 480.
Every face that existed on both devices agreed on who it was, and the
people rows were identical — the merge was fine. The missing 3,170
confirmations were on faces the tablet did not hold at all: the
desktop's 16,080 faces from the original detector, on 4,310 images,
detected before schema V14 kept the quality reading.
Those faces were in shards the tablet had already downloaded, in
August's export. `import_from_shards` looked at them on every sync pass
and declined each one, because a face without a quality reading was
"work this device cannot finish": adopting it would write the run
marker, and the marker was what stopped an image being looked at again.
That was true when it was written and has not been since the quality
repair existed — that pass lists its work by `f.quality IS NULL`, not by
the marker, exactly as the eye pass does, and faces without an eye
reading were already adopted on that reasoning.
The refusal had no exit. V14 had deleted the markers of every image
holding such faces so the quality pass would find them, and
`export_to_shards` walks the markers, so the desktop never re-exported
them either; the unmeasured August copies were the only ones there
would ever be. The tablet's answer was to queue all 17,727 images for a
re-detection of its own, a fetch of the whole library, while holding
the faces on disk.
Adopt them. The receiving device's quality pass measures them when it
reaches them, and the desktop's confirmations match onto them by box
overlap on the next catalog merge. The test that asserted the refusal
now asserts the adoption and that the image is still owed to the pass.
Eighteen declared operations, not fifteen; JPEG XL is an export format;
the grid does not filter by keyword, only the catalog's query can; and a
panorama's provenance is a sidecar beside the composite, not a history
step in it.
It said 0.9.0 against a 0.13.1 tree, listed focus peaking and burst
grouping as unbuilt when both have shipped, and opened with a page of
prose about the display path before saying what the application does.
Lead with what it is and a picture of it, say how to get it on each
platform and what state each channel is in, keep the honest account of
what is missing, and put the manual first in the documentation table.
A CLAUDE.md at the root, for anyone changing this code: the ways a redraw
and a catalog read came to cost half a second per click, what each fix
looked like, and how to measure the next one against a copy of a real
catalog.
"How many images still owe a quality reading" was a correlated EXISTS per
image over `faces`, and the face row is 8 KB of embedding and crop before
the column it looks at, so each count opened every row. Six such counts
run on every open of the Identity screen and at the end of every sweep:
160 ms on the reference library.
V19 adds three partial indexes holding only the faces still owing each
pass, keyed on the image and carrying the model id the predicate reads,
and replaces `faces_image` with `(image_id, model_id)` so "does this image
hold this embedder's faces" is answered from the index too. The planner
takes a partial index when the count is driven from `faces` and ignores it
inside the EXISTS, so `Needs::Face` carries the per-face fragment and
`repairs::count` spells the query from the faces' side; the list and the
per-image check keep the EXISTS. A test holds the two spellings to the
same answer for every repair.
Every `move_to` guaranteed its destination's parent with a `MKCOL` for
each ancestor down from the account root, and a trash folder under a
library root several levels deep meant three round trips answering
`405 Method Not Allowed` before the one `MOVE` that did anything — for
every image of a delete, on a connection built for that job.
The backend now records the collections it has confirmed exist and asks
about each once. It lives for one job, so a folder another client removes
mid-batch is the one case this misses, and the `MOVE` then reports the
`409` rather than hiding it.
`faces::people` grouped `face_person` after a LEFT JOIN over every person
and sorted the lot by name; the rail then discarded the empty, unnamed
groups a regrouping pass leaves behind — 17,000 of 19,000 rows on the
reference library. `people_in_use` filters them in the WHERE and joins
`people` to face counts aggregated first (2,000 groups), so the sort sees
only the rows that will be drawn. `count_unassigned` replaces fetching
2,400 ids to take their length. `load_people` 22 ms → 10 ms.
`confirm_all` called `faces::confirm` per face, and `split_off` called
`reject` then `confirm` per face: each opens and commits its own
transaction, so a click on a group of several hundred was several hundred
commits. `faces::confirm_all` is two statements — clear the rejections the
confirmations override, then flip the rows — and `faces::reassign` does a
split's reject-and-confirm for every face under one commit. 16 ms → 2 ms
and 22 ms → 4 ms on the largest group.
A confirm or a reject changes one row and redraws the whole grid, and the
redraw re-read every crop blob of the selected person (4 MB for the
largest) and decoded every one — 316 ms per click on the reference
library's 754-face person, to arrive at the pixels already on screen.
`load_faces` now takes the crops the previous load decoded, keyed by face,
and moves each into its new cell; the blob read is skipped when every face
is already in hand. `refresh` drains the old cells into it rather than
cloning them. The redraw is 2.6 ms.
Every click on the Identity screen's face grid — confirm, reject, split,
rename, merge — redrew the whole screen, and the redraw recomputed the
coverage line. That line lists every repair's outstanding images to count
them: six scans of the images table with a correlated EXISTS over the
8 KB face rows, an ORDER BY the job's visiting order, a Target with its
path per row, and a thumbnail-index query per image with faces. On the
reference library (24k images, 19k faces) that was ~200 ms of the
~540 ms each click cost, spent computing a figure a confirm cannot change.
`refresh` now takes what changed: `Changed::Identities` re-reads the rail
and the grid and leaves the coverage line alone; `Changed::Library` — an
open, a sweep ending or stopped, the face data deleted — re-reads it too.
For the times it does run, `repairs::counts` counts instead of building
and dropping the lists, and the thumbnail store is read once
(`ThumbStore::held`) rather than probed once per image in the audit, the
outstanding list and the proxy repair.
`identity_bench` is the measurement: the reads a click performs and the
batch writes, timed against a copy of a real catalog.
docs/manual/README.md is a tour for a photographer opening DarkRoom for
the first time — one picture per thing, moving where movement is the
point. tools/manual/ is how the pictures are made: drive.py puppeteers the
desktop build on a private Xvfb (launch, click, drag, type, screenshot,
record), scenes.py is each picture as a script, and record.sh runs them
all over a folder and writes the results into docs/manual/media/.
The media is in LFS, with the CI pulls excluding it as they exclude the
fixtures; a screenshot changes wholesale when the interface does.
Nothing in the pictures shows a person, by design: the demo library is
seventy urban and alpine frames, chosen from the catalog's rows that face
detection found nobody in.
The traceability matrix is regenerated here after the rebase that
brought this branch up to master.
Without focus the keys typed into the keywords sheet went to the grid
behind the scrim, and the Return meant for the keyword opened a
photograph. The naming sheet already takes focus on open; do the same.
An empty trash said "No images found — check the library folder and which
formats are ticked", which sends someone off to fix a library that is
fine.
An empty device destination read "Ask each time" on the settings page,
and nothing asks: an export made with the field blank is refused with
"no export folder is set". Say what will happen.
wgpu reports a device out of memory by panicking, and a twelve-frame
merge on a GPU another process is using is where that happens. The panic
unwound the worker, the sender went with it, and the page sat on "Stop"
with every control disabled and nothing to say why — the crash record on
disk was the only sign. Catch the panic and send it as a failure, and
treat a closed channel with no final event as a dead worker too.
The eyes are per layer and outlive the mode, so a photographer coming
back finds the layers they were looking at still lit. But the tint is a
way of looking at a mask, and outside Local there is no mask being looked
at: the sky stayed red through Repair and back in Photo, a mode that had
been left leaving its overlay behind — the fault ui-navigation.md D-N1
exists to prevent.
Five chips in one row declare 440px, and the develop column takes the
widest panel's request — so selecting a category mask levered the sidebar
past the window's edge, clipping the histogram, the group strip and the
subject list. The same trap ChipGrid's comment records for film formats.
A heading was drawn only on a cell that both began a month and began a
row, so at seven columns most months were never named, and the one
heading on screen — always on the window's first cell — was wrong about
every row below it. Worse, two headings drawn on the same cell overprinted
each other. Now a row carries a heading whenever its first cell's month is
not the one last announced: a month starting mid-row is named on the next
row it opens, one row late and right about everything under it.
Every shard is in WAL mode and every put opens its own connection, so
while thumbnails are being generated on several threads — which is when
the first sync pass runs — the log is never checkpointed and the main
file holds whatever the last quiet moment left in it. For a shard created
seconds earlier that is nothing: a zero-byte file with the schema still
in the log. The sync read that file and uploaded it, and every other
device merging it failed with "no such table: thumbs" on every pass.
Copy the shard through SQLite's backup API into scratch first, which
serialises against writers and carries the log, and upload that.
Confirming "/" in the folder picker set an empty root, which the launch
model read as no root at all: "Open library" stayed disabled after the
question had plainly been answered, and a folder library — whose folder
is the whole library — could never be opened without first descending
into a subfolder of it. The empty string was carrying two meanings.
Record the choice as its own fact on the account (`root_chosen`, defaulted
so existing configuration loads unchanged), treat a folder endpoint as
chosen by definition, and let the launch screen say so: a folder is shown
as a LIBRARY rather than an ACCOUNT, the second question becomes an
optional "scan only a subfolder", and the library header names the folder
instead of calling it "· whole account".
The scan reads the first 256 KB of a file for its metadata. A camera
writes its IFDs at the front, so that is the whole structure; the linear
DNG a merge writes puts its first IFD after the pixels, and rawler,
given the head alone, finds no decoder in it. The composite was
catalogued without a date and sorted to the very end of the grid, after
every dated photograph — which is where a panorama merged on the tablet
went unfound.
dr-decode's own TIFF reader now reads through a head and a tail at a
known offset; trailing_ifd says where the tail starts and
metadata_split reads the two together. The scan, when the head fails
and points beyond itself, fetches from the IFD to the end — kilobytes —
and dates the file from both. Tested against the writer's own output.
The Malvar "R at green in R row" kernel weights the two greens two
sites away along the row at -1 and the pair up and down the column at
+1/2. The shader had the two swapped, in the comment as well as the
code, so the transcription checked against itself. Both sum to zero
and reconstruct a flat patch exactly, which is all the tests fed it.
On an edge the correction at green sites is half strength and the
false colour doubles: 0.375 against 0.19 on a grey step, and a
blue/yellow zipper around every clipped highlight at 1:1. The other
three kernels and the CFA tables were right.
A grey vertical step now runs through the pass; the transposed kernel
fails it at 0.375.
The reference desktop's only system ONNX Runtime is Arch's
onnxruntime-opt-cuda: 1.29, built without TensorRT and against cuDNN 8
on a cuDNN 9 machine. The probe rejects both providers correctly and
the app runs on the CPU provider, which is right and not what anyone
wants. runtime/ beside the models is now searched ahead of /usr/lib,
tools/fetch-desktop-runtime.sh fills it with the four libraries from
the current onnxruntime-gpu wheel (cuDNN 9, TensorRT 10), and the
About caption lists every rung that lost and why, not only the first.
Verified: the app selects TensorRT from that directory with no
environment variable set.
The unpack list gained migan-512.onnx without its length following;
nothing on the desktop compiles that crate, and the first Android build
of 0.13.0 stopped there.
A fill that went wrong took a seven-minute merge to look at again. Now
DR_FILL_DUMP=dir makes the merge write what the filler was given, and
the fill example runs fill_border on that, or a crop of it, on the engine
and writes coarse, each band and the feathered result as PPMs — seconds
per attempt on TensorRT. Both examples take DARKROOM_ORT_DIR as the app
does, and --wait-engines lets a compiling rung finish before timing.
A Border choice beside the projection — crop to the picture, or fill it
— that redraws the preview filled so the invented pixels are seen before
they are confirmed (FR-MRG-1), greyed with the reason when the model is
not there. The job fills at half the composite's resolution in a
display-ish space (white balance, matrix, gamma; invertible) and samples
the result back into the linear DNG wherever no frame reached; the
sidecar's merge line says border filled and with which knobs.
Experimental because the fill is right in thin borders and wrong in deep
corners, where the model's Places2 prior puts clouds in sky and water
under grass; so its six knobs — working scale, edge erosion, coarse pass,
band width, mirror depth, seam feather — are sliders under the choice,
each committing a redraw, until the defaults are right.
dr_pano::fill owns everything the model does not — which tiles, what
context, how to blend — behind an Inpainter trait, and dr_pano::migan is
that trait over the shipped generator on the inference engine.
The known content is mirrored across the coverage edge into the hole and
a 256-px ring, the nearest 48 px folded, so the model interpolates between
real and mirrored sky rather than extrapolating into nothing. A coarse
pass at a quarter decides the structure with the whole border in a few
tiles; fine passes in 96-px bands from the edge outward texture it; the
seam is blended over a feather inside the real edge. Every knob is a
Params field, and an Observer hears each stage for whoever is looking at
why a fill went wrong.
A border fill acquires the filler once a tile, and each acquire hashed
the 28 MB model twice — 60 ms a tile, a third of the tile's run on a
throttled TensorRT. The Model keeps its hash from open.
Sargsyan et al., ICCV 2023; MIT code and weights (models/LICENCE.md),
exported by tools/export-migan.sh at a fixed 1×4×512×512 from the
authors' checkpoint — six operator types, 28 MB, in LFS like the rest.
The package installs it beside the scene model and the APK unpacks it
with the others.
The fused shader stored black with alpha 1 for a pixel whose source
coordinate left the frame, and the merge's warp averaged it in like any
other: a dark, badly interpolated fringe along every frame's edge, visible
as a seam wherever a frame ended and, later, as the edge the border fill
continued. The display keeps its opaque black; CameraLinear stores alpha 0
and the warp weights each sample by the alpha it interpolated, dropping a
sample that has none.
MI-GAN is plain convolutions, so every rung serves it and none needs a
special form; the role exists so resolve_model and the probe's fingerprint
know the model, and so the merge job can open it through the engine rather
than tract, which takes 7.4 s a tile for it.
A library's records are never all complete at once. A face found before
its quality was kept has no quality; one found before the eye models
existed has no reading; one adopted from a peer's shard has no crop; an
image the fast detector examined on a 1024 px proxy has boxes the current
detector would not have drawn; an image the scan stat'ed has no capture
date. On the reference library that is 17,762 faces under the bare
w600k_mbf id with no quality, no reading and no dense landmarks, 4,144 of
them without a crop, beside 12,217 images the fast detector examined and
found nothing in. Every one of those gaps was its own pass — V14's
measuring pass, §17.5's eye pass, the sweep's proxy repair, the sweep's
detector upgrade — with its own work list, its own count and its own idea
of done, and adding a per-face field meant adding a pass. There was no
pass at all for the case the library is actually in: boxes and landmarks
drawn by a weaker detector on a proxy, which every later per-face pass
would have read from.
dr_ui::repairs replaces them with one job over a registry. A Repair names
one thing a record can lack — the predicate that says which images still
owe it, the input its handler needs (a header, the original, or a native
render), the handler, and what to record for an image that can never be
done. The job unions the predicates into one work list, fetches each
image once at the most any claimant asks for, renders it at most once,
and runs every handler whose predicate that image still matches, checked
again before each because a detection writes every field a per-face
handler would fill. The registry today: face-proxy, face-quality,
face-eyes, face-crop, face-detection, face-upgrade, metadata — the last
there to say that this is not a face job. Adding a field is one entry.
A repair's predicate is the only definition of its work: the count the
settings page shows, the list the job fetches and the check before its
handler run are one predicate, so the job converges. That is why the
registry is cut to what the device can do rather than listing what it
skips — an entry is a count and a set of originals to fetch — and why an
eye reading that cannot be cut is not a criterion.
The catalog side is generic to match: record_updates writes whichever
fields a FaceUpdate carries and re-marks the image so the shards export
it; faces_needing and count_needing answer a predicate the caller
supplies, replacing the measuring pass's three special cases.
Two buttons on the settings page run the job and differ in one
predicate. "Index faces" converges on coverage: has anything examined
this image. "Re-index every face" converges on provenance: face-detection
claims every image with no marker under the chosen detector, in either
of its forms (FaceDetector::model_ids, so a desktop in f32 and a tablet
on the Hexagon do not re-index each other's work), and a marker saying a
weaker one looked is not that. An original over the fetch budget is left
exactly as it was under the re-index, where the sweep marks it examined:
a re-detection with nothing found would delete the faces, and "cannot
fetch" is not "no faces".
record_detections replaces an image's faces and carried only the user's
confirmations onto the new ones, by box overlap above 0.5 IoU. Everything
else on the old faces was dropped: the suggestions the last grouping pass
made, and the people the user had said a face was not. On the reference
library that is 13,011 suggestions and 77 rejections beside 3,778
confirmations — a re-detection of it would have been correct by
FR-CULL-12's letter, since suggestions are derived data, and would have
handed back a People screen of strangers.
Now every old face is read before the delete — box, vector, assignment,
rejections — and matched to the new faces one-to-one, best pair first. A
pair qualifies when the boxes overlap at all and either the overlap alone
says so (IoU above 0.5, the old rule) or the embeddings do (cosine above
SAME_FACE_COSINE, 0.45, the reference library's P≈0.95 line). The
embedding route claims the box a low-resolution pass drew badly enough
that overlap alone would not; the vector is also what breaks the tie in a
group photograph, where two neighbouring faces overlap both new boxes.
Overlap is required on both routes, because the same vector elsewhere in
the frame — a mirror, a print on the wall — is not the same face and must
not take its name. Onto the matched face go the assignment as it was,
confirmed or suggested with its probability, and every rejection.
The merge's match_faces still matches by overlap alone across devices; it
is the same question and is not changed here.
ONNX Runtime's default is already its fullest level. ort-tract maps any
level but disabled to tract's optimiser, whose slice pass divides by
zero inside the segmenter's graph — a panic across the C API and so an
abort, which is what stopped dr-ui's develop test. The app never asked
tract for that and does not start now.
On the tablet the engine compiled arcface for the NPU: the routing
compared the form a rung wants with the form on offer, and for the
embedder both are f32, so nothing said no. A rung now says which roles
it serves at all, and the Hexagon does not serve the embedder (§7 —
its vectors must compare across devices). Tested at the routing seam.
dr-segment's onnx_probe example still named ort-tract, which is what
stopped the workspace test build.
The strict flag refused the Hexagon over the ten quantise/dequantise
nodes at the graph's edges that QNN declines by policy, which cost
microseconds. A provider that hands real work to the CPU is slower than
the CPU floor and the timing already rejects it; the tablet measured
2.3 ms on the NPU against a 29.7 ms floor.
ONNX Runtime's errors open with a source path and a template signature;
the first 160 characters of a CUDA failure were all signature. The
reason now starts at the first word a person can act on.
XFeat's two exports are a Keypoints role now; the crate no longer names
tract, and the app compiles TensorRT engines for both ahead of the
first merge. The probe picks the smallest *detector* rather than the
smallest file: the tablet's first run chose the 112 KB eye classifier,
which has no int8 form, and reported the Hexagon as failed for want of
one.
The first int8 files found no faces at all, and for two reasons the
tool now guards against. The calibration set was landscape photographs
with no faces in them, so the score head's ranges had never seen the
face regime; the set is now proxies from the library itself. And ONNX
Runtime's strided and moving-average calibration modes both degrade
these graphs measurably (a quarter of the faces at eight images, none
at ninety-six), while driving the calibrator in chunks by hand gives
ranges identical to a single pass — so the tool does that, four images
at a time, and feeds quantize_static through its range cache.
Measured against f32 over 400 proxies (docs/inference.md §10.1): the
10g form finds every face above 32 px the f32 form finds; 500m and
2.5g find 96%, and what they lose sits at a median confidence of 0.52
against the 0.50 threshold. Shipped with the number on record.
The Android unpack list gains the three int8 files; without that the
tablet never saw them. D13's runtime half records the reopening.
tools/quantise-models.sh writes the QDQ form QNN's HTP backend takes
whole: opset 17, per-channel int8 weights, uint8 activations, ranges
from running the f32 graph over photographs fed exactly as the app
feeds them. The calibration is strided, four images at a time, because
every ONNX Runtime calibrator holds each image's whole set of
activations until it folds them — a gigabyte an image on the 10g
detector, and an OOM kill with no message when folded once at the end.
Release-time, never on the device (docs/inference.md §5): it needs
real photographs and a person reading the recall measurement that
gates whether each file is offered.
The desktop names where a package may have put libonnxruntime — an
override variable, beside the executable, the package's own library
directory, the Flatpak prefix, the system library directory — and
Android points at the APK's native library directory, which is also
what Qualcomm's DSP loader must be told for the Hexagon skel. Android
starts the engine at the end of the model unpack rather than at launch,
because the probe fingerprints the model files and a first launch has
none until then.
The About panel gains an Inference row beside Graphics, re-read every
two seconds while the probe runs and engines land, and faces.model_id
carries the detector's form: an int8 detector finds a different set of
faces and is a different population (docs/inference.md §7). A
low-memory signal drops every idle session with the GPU caches.
The APK assembly bundles ONNX Runtime and the Qualcomm HTP libraries
from Maven, fetched by tools/fetch-android-runtime.sh with their
published checksums; RUNTIME_DIR=none builds the tract-only APK, which
is a slower app and not a broken one. The desktop packages carry no
runtime yet.
Two probe fixes from the first desktop run: the floor must not be
built with CPU fallback disabled, and a versioned libonnxruntime.so is
a runtime too. On the reference desktop the probe now loads ONNX
Runtime 1.30, measures 30 ms on the CPU provider, and selects TensorRT
at 1.5 ms.
One crate names the runtime, the providers and the devices; dr-face and
dr-segment ask it for a session by role. It hands ort an API table once
per process — from a libonnxruntime it dlopens when the app names a
directory holding one, otherwise from tract — so the Rust build stays
free of C on every target and a package can install the runtime as a
file (docs/inference.md §3).
Sessions live in a registry behind a Model handle that holds the bytes,
not the session: every use refreshes a timestamp and a reaper unloads
whatever sat idle past the decay. A scan that runs the detector on each
image never lets it go idle; a click in the develop view lets the
segmenter go after thirty seconds; a handle used after that reloads,
and reloads on a higher rung if a compiled engine has landed meanwhile.
The probe walks the platform's ladder by building strict sessions and
timing them against the CPU provider, caches the choice against a
fingerprint of the runtime, driver, hardware and models, and compiles
engines for the selected rung in the background, smallest model first.
Nothing in this commit turns the native path on: the apps still run on
tract until they call init with a runtime directory.
tract runs every model on one core on every platform. Measured against
ONNX Runtime's providers on the MagicPad 2 and the reference desktop:
ORT CPU alone is 3-10x, the Hexagon at int8 runs the detectors in
1-3 ms, TensorRT is ~2x the CUDA provider. NNAPI, XNNPACK, WebGPU and
CUDA int8 were tried and excluded with the numbers that excluded them.
The spec keeps the build C-free: ort::set_api takes a table from a
dlopened runtime or from ort-tract, chosen once per process. Rungs
are chosen by building a real session, cached until an input changes,
and compiled engines are built in the background after the first
frame. The embedder stays f32 everywhere; int8 detectors are a
distinct model_id and are gated on a recall measurement.
Read and measured, not built. The bare 512 generator exports at a fixed
shape and loads under tract with nothing unsupported; at f32 on the
desktop CPU it takes 7.4 s per 512×512 tile, which puts a full-resolution
fill of the fixture's border at ten minutes. The three routes that would
make it viable are recorded, with the quarter-resolution fill the cheapest
and Hexagon int8 the one the model was designed for.
Picking a chip stored the choice for the merge and changed nothing on
screen — the chip did not even highlight, since the selected property
was never written back. Now the pick is reflected, and the job, waiting
for its decision, takes a Preview request, draws the alignment on the
chosen surface at proxy cost and reports again; the drain puts the new
picture and its size up. Auto is the surface the field of view suggests.
Also:
The largest rectangle inside the frames' coverage is found a row at a
time — a histogram of consecutive covered rows and a stack pass per row —
so the composite is never held to be measured (FR-MRG-11). It is written
as DefaultCropOrigin/DefaultCropSize (FR-MRG-4): the file opens on the
picture, the border is still in it, and resetting the crop shows it.
rawler reports the crop as the picture, which the test checks.
FR-MRG-4 records the question raised the same day — fill the border
rather than crop it — as open: a non-generative fill through the heal,
or a generative inpainter with its licence and weights. Neither decided.
A sweep whose frames overlap by more than half reads as one photograph,
and the page's job is to show frames. Each footprint is walked along its
border and drawn in amber where it lands, so twelve frames look like
twelve and a misplaced one is visible as such. The preview may take most
of the page's height rather than 320 px.
derived_from and merge are top-level sidecar fields (FR-MRG-6): one line
per source in order, and how the composite was made. A build that
predates them keeps the lines as unknown and writes them back. The job
writes the sidecar beside the composite and stages it with its own
record when the composite goes through the outbox.
DARKROOM_START_MERGE=a.CR2,b.CR2 lands on the merge page at startup with
the job running on local files, on the model of DARKROOM_START_IDENTITY,
for looking at the page where synthetic clicks do not reach it. The fetch
and the start are shared with the grid's button.
panorama.md §11 records what exists, the fixture's figures, and the six
things still open, auto-crop first.
dr_ui::merge is the orchestration with no interface in it: decode each
frame to sensor data and build its graph as a session would (orientation,
lens profile); render each through the camera-space tap at proxy size and
detect keypoints there, so the alignment is measured in the undistorted
frame the tiles are rendered in; align; solve one gain per frame from the
proxies' overlaps; draw the aligned set in colour for the page; then wait.
Nothing is written until a Decision arrives (FR-MRG-1). The merge writes
a linear DNG through the outbox with a destination record, so the drain
puts it beside its sources on a folder library and a server alike, and
the library rescans (FR-MRG-3).
merge.slint is the page, on the import page's model: the alignment
table with a failed frame named on its row and the button held off
(FR-MRG-5), the preview, the projection choice, Stop and Back. A
"Merge to panorama" button joins the grid's selection bar at two frames.
Headless, the example produces the fixture's 22 993 x 5 980 DNG in 45 s
on the reference desktop, exposures balanced across the stop of drift.
merge.wgsl warps one camera-space tile into one output chunk — output
pixel to direction (the projection maths of dr_pano::projection, verbatim),
direction to the frame's camera, camera to source pixel, bilinear by hand
from four textureLoads because rgba32float is not filterable — and adds it
into a storage-buffer accumulator weighted by its distance from the
frame's edge. A resolve pass divides by the weights and packs sixteen-bit
samples at the sensor's scale with a coverage bit.
MergePass::merge drives it: bands of rows, chunks across a band, and for
each chunk only the frames whose footprint meets it, each rendered as the
source rectangle the chunk needs and nothing more. The working set is one
chunk, one tile and one band (FR-MRG-11); the frame textures are the
caller's to cache. Feathered, not seamed; gain a scalar per frame — the
blend quality is panorama.md §10's step 5, after the path writes a file.
dr-export gains write_linear_dng — LinearRaw, DNG 1.4, u16 samples at
the sensor's scale, the body's matrices with their illuminants, the
as-shot neutral, the EXIF block an export writes — streamed strip by
strip through a closure so the composite is never held (FR-MRG-11). The
tiff crate's directory is a map, so PhotometricInterpretation is written
over what new_image set, which is the trick the S15.1 spike thought it
had to hand-roll around. The test reads the file back through rawler.
dr-decode's RawImage carries samples_per_pixel (a linear DNG is 3), the
body's profile with its calibrations mapped back to EXIF illuminant
codes, and the cleaned make and model. The GPU uploads a three-sample
image as it is, normalised by black and white like a photosite, through
a full f16 conversion — subnormals kept, because a 14-bit LSB sits at
f16's smallest normal and rounding it to zero would crush exactly the
shadows the file was written to keep.
compose_camera_linear composes the fused pass with an empty operation
list, the file's orientation as the baseline, a view rect for the tile,
and a store of rgba32float. On the GPU, render_camera_linear is the only
entry that accepts it: it fills the profile uniforms neutral — unit white
balance, identity matrix, curve off — so what lands in the texture is the
sensor's numbers after the lens warp and nothing else (FR-MRG-2). A third
bind-group layout carries the format, as the linear one does, and the
readback is generalised to any pixel width for the f32 copy.
Thirty-two bits because the composite is written back at the sensor's
scale: a 14-bit sensor has 16 384 steps to white and f16 keeps 2 048 of
them in the top octave.
Twelve real frames from the fixture set now align in 4.5 s — 4.4 s of
matching, 118 ms of bundle adjustment — where the first run took 51 s and
left the first two frames out.
The matcher computes each pair's similarity matrix once, across the
cores, with a dot product written to vectorise; both nearest-neighbour
directions read it. The frames that failed were portrait: fitted into the
landscape input they used 512 of 1024 px, and their thin overlap did not
survive at half resolution. The same weights are now exported at 768×1024
as well and the detector picks the shape by aspect. The example aligns
from embedded previews and draws the set on a cylinder; on the fixture the
sweep is 152° at a fitted 47.9 mm against the EXIF's 50, RMS 1.5 px, and
the overlaps show no ghosting.
A new crate holding the CPU half of a merge (FR-MRG-10): the grayscale
proxy with orientation, the XFeat decoder ported step for step from the
reference detectAndCompute, mutual-nearest-neighbour matching, a robust
pairwise homography with the focal length read off it, a hand-rolled
Levenberg–Marquardt bundle adjustment over every rotation and the focal,
the three output projections, and align(), which chains it all and names
the frames it could not place rather than guessing (FR-MRG-5).
Dependency-free without the xfeat feature — linalg.rs says why the dense
algebra is hand-rolled — and tested on synthetic sweeps whose answer is
known exactly. The noise test records the single-row degeneracy: one
pixel of noise is a tenth of a percent of focal, which is a uniform
stretch of the sweep, not a misalignment.
tools/onnx-probe-on-device.sh cross-builds dr-segment's onnx_probe
without the embedded segmentation model, pushes it with a model to the
attached device and times two runs. The 768×1024 XFeat export takes
~400 ms on the reference tablet's NEON cores against ~300 ms on the
desktop, with identical output ranges — inside NFR-MRG-1's 1 s per frame.
The blend half of S15.4 waits for a chunked blend to exist.
fixtures/pano/2025-08-05: _MG_8320 … 8331, one portrait hand-held sweep
at 50 mm with a stop of shutter drift and sky in every frame — the set
§3.11 is built against, with each of those facts named as the test it
is. fixtures/** is tracked in LFS like the models but with the opposite
default: CI's pulls exclude it, so a build never fetches 325 MB it does
not use.
Camera-linear u16 samples on the first source's black-subtracted scale
with its white level, never rescaled to fill 16 bits, with its body,
matrices, illuminants and as-shot neutral carried — so the panorama is
developed afterwards as one photograph from the sensor's own numbers.
The only thing a warp cannot preserve is the colour filter array, and
the clause says so.
The fused chain, as operation.rs's tests fix it, is warp → as-shot white
balance → operations → base curve → camera matrix → store. LinearWorking
stores after the matrix, so the existing linear tap carries the body's
base curve, and a composite stitched from it and developed as an
unprofiled body would render that curve twice.
FR-MRG-2 therefore stitches camera-linear RGB — after the warp, before
white balance, curve and matrix — and the composite carries the first
source's body, matrices and as-shot neutral so its own develop applies
the profile once. The composer already makes this a uniform question:
white balance, matrix and the curve flag are reserved uniforms, so the
tap is a compose entry with no operations and a render entry that fills
them neutral. panorama.md §5.1 states the shape and asks for f32 buffers.
tools/export-xfeat.sh exports the convolutional network alone at 768×1024
grayscale, on the pattern of export-seg-model.sh: thirteen standard
operator types, no dynamic axes, the keypoint decoding left to Rust.
examples/onnx_probe loads it through the ort-over-tract backend the app
ships with nothing unsupported and runs it in ~300 ms on the desktop CPU.
The weights are Apache-2.0, read from the repository's LICENSE, with no
grant on the checkpoint — recorded in models/LICENCE.md before they land,
as FR-MRG-8 asks. The probe stays: the next model will need the same
check.
A hand-rolled 64×48 LinearRaw DNG — one IFD, 16-bit RGB, DNGVersion,
ColorMatrix1, AsShotNeutral — comes back through rawler 0.7 with cpp 3,
the samples in the order written and the matrix parsed into the camera
definition; CameraProfile::extract builds a profile from it. ImageMagick
reads the same bytes.
dr_decode::decode currently accepts the file as CFA and passes three
times the samples on, so the cpp == 3 branch is the decode work FR-MRG-3
needs, and the only decode work. panorama.md §8 records the result.
A merge writes a new source file beside its sources (D18) rather than a
multi-source Version, which answers the schema question §7 had been holding
open for panorama, HDR merge and focus stacking together. The panorama is
undeferred as FR-MRG-1 … 11; the other two stay in §7 with their data model
decided.
FR-MRG-10 and 11 fix where the work runs — every per-pixel stage on the GPU,
the composite never held as one texture — because the output exceeds
max_texture_dimension_2d before it exceeds memory. panorama.md carries the
stage table, the chunked output driver, the model licences and the porting
sources. S15 gates all of it.
Coverage falls from 83.0% to 77.2%: thirteen requirements entered with no
code, and outstanding.md §11 says so.
Once every image has been through the detector and only readings are
left — the state an already-indexed library is in the day the eye models
arrive — the button reads "Read eye state" rather than promising to
index, and the coverage line says what the faces are waiting for.
The 106 points the eye boxes were cut from, stored beside the reading as
16-bit fixed point over the frame: 424 bytes a face, a seventh of a pixel
on a 6000-pixel frame, where f16 at the same size would have been six.
Derived data like the embedding, kept for the same reason — it cost a
fetch and a model run, and the next per-face pass should run from the
catalog. Shards carry it; a peer's shard from before it is still read.
The register grew both clauses the same day this was built: FR-CULL-8a is
the per-face state the reading is, and FR-CULL-13 is the rule that a
signal is shown and filtered and never writes a judgement. The tags,
faces.md §17 and catalog.md now say which is which; FR-CULL-8a records
what of it is built, and that its third model is under the InsightFace
grant by the same decision as the pair.
The people filter was served from faces_image without touching a row;
reading the eye columns in the same subquery touched every one, and
ALTER TABLE had put those seven floats after the embedding and the crop
blob. One count took 24 seconds on the reference library, thirteen of
them system time. faces_eyes covers the subquery again: five
milliseconds.
FR-CULL-13, with §3.9.1's exclusion of blink detection re-read as the
exclusion of blink selection it always was: the stored fact and the chip
are built, a pass that picks the frame where everyone's eyes are open is
not. faces.md §17 has the models, the crop measurements, the four-state
rule and its floors, the native and proxy sheets read face by face, and
what remains to measure.
An "Eyes open" chip beside the people chips, offered only while someone
is chosen and dropped when the last person goes, so no term narrows the
grid with nothing on the bar to say so. It compiles the rule in
dr_face::eyes into the person's face subquery — Anna, eyes open, whoever
else is blinking beside her — and drops a frame only on a closed eye that
could be read: sunglasses, eyes too small or soft to read, and faces never
read all pass, so an old library shows everything under the chip until
the measuring pass has run. A test drives the same readings through the
SQL and through the rule and requires them to agree.
The People screen badges a face "Eyes closed", "Sunglasses" or "Eyes
unclear" so the reason a frame is or is not in the grid can be read off
the face; the sweep loads the three models when they are beside the pair
and reads eyes on the indexing and measuring passes from the native
render; the coverage line counts unread faces as work to measure so an
already-indexed library keeps its Index button. The term travels with the
place.
2d106det for the eye contours, OCEC for open or closed, SGC for
sunglasses — all three pinned to a batch of one by the same script as
the pair, and installed by every packager so the eyes-open filter works
out of the box. The two classifiers are MIT, code and weights; the
README records their provenance, SGC's undocumented training set, and
the hashes as fetched and as shipped.
Per eye P(open), the pixels across its box and the sharpness of the
patch; and P(sunglasses). The verdict — open, closed, sunglasses,
unclear — stays a rule in dr_face::eyes so the floors can move without
re-measuring twenty thousand faces. Shards carry the same seven, and a
peer's shard from before any of them is still read.
SCRFD's eye point places a face, not an eye: on turned and smiling heads
the classifier's window had the eye in a corner, and two model-free ways
of re-centring it — the darkest blob, the most contrasty window — both
lost open eyes (19 → 15 and 19 → 9 of 25). Three landmark models were
then run over the same faces; Face Mesh V2 and InsightFace's 2d106det
tied at 22 of 25 and 2d106det ships, being the cheapest by far and under
the grant the detector and embedder already carry. The eye box is the
tight bounding box of its ten lid points, cut upright from the native
render, which is what the classifier was trained on.
The larger change is that the reading now carries, per eye, the source
pixels across the box and the sharpness of the patch — because the
commonest wrong answer on the reference library was a soft eye read as
closed, and a classifier shown a smear will always say something. An eye
under either floor, or narrower than six tenths of its partner (the far
eye of a turned head, whose contour collapses), is not asked; a face with
no readable eye is a fourth state, Unreadable, that no filter drops. On
twenty native renders the one real blink is caught, the laughing faces
are closed, the profiles are judged on the near eye, and the one thing
left beyond any floor is a face with a pot held over it.
Three nullable columns beside quality — P(open) for each eye and
P(sunglasses) — because the verdict is a rule with thresholds in it and a
rule belongs in code, not in rows that would have to be re-measured. NULL
is "never read": a face from before the models, or from a device without
them, and every reader treats it as unknown rather than as closed.
The measuring pass V14 built for the embedding's length is what fills
them, so the sweep's work list now also names faces with no eye reading
— but only on a device that has the models, or it would fetch every
original to do nothing to it. A peer's shard without the reading is still
adopted, unlike one without the quality: the pass finds this work by the
NULL rather than by the run marker, so adoption costs it nothing.
Two MIT classifiers from the same author as the reference pipeline's
whole-body detector: OCEC answers P(open) for one 40×24 eye, SGC
P(sunglasses) for a 48×48 head. Both load in tract once their batch
dimension is pinned by tools/fix-face-model-shapes.sh, like the embedder.
The crops come through the same fitted similarity the aligned face does,
so an eye window is a constant in template units rather than a second
warp, and a tilted head yields an upright eye. Measured on 60 proxies
from the reference library: the eye window plateaus at 22×11, the S
variant beats M and L (which overfit their own domain), and for
sunglasses the aligned face beats a head framing but the higher of the
two catches 11 of 12 pairs against 9 for either alone.
The reading keeps both eyes and the sunglasses number apart, because a
wink averages to the least informative value and a lens of dark glass
draws a confident answer from the eye classifier — over a woman in
sunglasses it read the right eye 0.97 open. Sunglasses take precedence,
and a face behind them is neither open nor a blink.
Its plugin section still asked for the contradiction to be resolved, its
render-path section still asked whether FR-DSP-2 was a requirement and
said NFR-RES-2 had no answer, and its closing section still called D12
open. Each now records what was decided and keeps the argument that was
weighed, so the document reads as the history it says it is rather than
as a plan the register has moved past.
dtourolleandClaude Opus 5