Detection now walks a browser-seekable H.264 proxy in MediaPipe's VIDEO
running mode. The PNG sequence it replaces was 112MB for 7.6 seconds at
1440x1920 and 1.1GB at the 900-frame limit; the proxy is 6MB, and landmarks
detected off decoded H.264 rather than off the PNGs moved at most 0.0033 of
frame width.
Three things had to be true for video mode to work, and each was measured
against the same footage decoded to PNGs:
/blob/<digest> answers byte ranges. Django's FileResponse does no Range
handling, and a media element handed 200 with no Accept-Ranges reports an
empty `seekable`, no-ops every currentTime write, and detects frame one
ninety times without raising.
A seek aims at the MIDDLE of its frame. Aiming at i/fps sits on a frame
boundary and landed one frame early 31 times in 91; (i + 0.5)/fps was exact
on all 91.
Timestamps are strictly increasing footage milliseconds. Video mode is a
tracker: a repeat leaves the graph in an error state every later call
re-throws, so the landmarker is discarded on failure, and passing the frame
index instead of i*1000/fps moved landmarks six times further from the
per-frame answer.
Frames are verified rather than trusted. requestVideoFrameCallback states
which frame it handed over, the walker discards any other and fails loudly
if the one it asked for never arrives — a stale presentation from the tail
of a previous seek is what produced "asked for frame 1 and it presented
frame 2" on a video whose seeks were in fact exact.
The proxy is re-encoded even when the upload is already H.264: HEVC is not
decodable everywhere, and footage identity is the proxy's digest. The JPEG
stills beside it are tracing references, outside the footage digest because
re-rendering them at another size is not different footage.
Verified end to end in a real browser against real footage: 228/228 frames
detected, a drawn roto face, 37 backend and 234 frontend tests green.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
There were two answers in the tree to "which stored bytes stop being valid when
this knob moves", and only one of them was checked.
`flow/address/block-knobs` is per block and asserted by biconditional —
`address-test` re-freezes the take once per knob and requires that the bytes
changed if and only if the key did. `domain/params`'s `:affects` was per area,
had no caller but a test asserting it returned what it was written as, and was
already wrong in both directions on the one entry where the two granularities
disagree: `:aperture-cut` claimed `#{:mouth}`, where it reaches no block, and
omitted the teeth, whose contour bytes it genuinely moves by gating
`condition/interior`'s smoothing. `:blink-cut` claimed `#{:eye}` and reaches no
block either, because a blink is `[:vis]` keys in tier 1.
So `:affects` and `affected-areas` are gone, and `address/knob-roles` is the
derived inverse of the table that is asserted — which is what a parameter panel
actually wants to ask. A knob absent from it invalidates no block, and that is
an answer rather than a gap.
Two new assertions keep the derivation from rotting at either edge: every role
in the table is reachable from some knob, and every knob a block declares is one
the registry defines. The second closes a real hole — `block-descriptor` checks
only that a knob was PASSED, and the freeze's `merge take/knobs` makes that true
of anything spelled like a keyword, so a typo in `block-knobs` would have named a
setting no slider can move.
228 CLJS tests, green.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Step 9. The tier split was the work; Django was the easy half.
Tier 1 — the authored scene — is the document, and it is addressed as
independently versioned leaves rather than saved whole, so one vertex drag
cannot clobber a collaborator's keying. `domain/leaf` is the document as
path -> value; `domain/wire` puts it on the wire as transit, because JSON
has neither integer map keys nor keywords and a save would quietly turn
`{0 v}` into `{"0" v}`.
Tier 2 — the dense channel blocks — is content-addressed by a hash over
every input, with the detector version inside every key through the
analysis the block descriptor names. `flow/address`'s `block-knobs` is the
invalidation table, and `address-test` does not trust it: it re-freezes the
take once per knob and asserts the biconditional, that a block's bytes
changed if and only if its key changed. That found `brow-pos` not depending
on `contour-avg` — the brow ring is smoothed, the raise is not.
Tier 3 — frames and audio — is served by the hash of its bytes out of the
same store. A manifest now names frames and carries a URL for each, so the
frame layout stopped being a shared secret between a shell script and a
ClojureScript namespace, and the `?v=` cache-buster went with it: a blob's
name is the hash of its contents, so a stale copy is not a thing that can
happen. The synthetic take's `audio.wav` moved to `static/arthur/` — an
asset the project owns, not an extraction that churns.
The server verifies rather than trusting a name it was handed: it
recomputes every key from the descriptor stored beside it, refuses an
analysis that declares no detector version, and refuses a document naming
blocks it does not hold. It hashes the descriptor TEXT, because JS prints
an integral double as `1` and Python as `1.0`, and a scheme where both ends
re-render the numbers disagrees on the first parameter that happens to be
whole.
Two loose ends from step 8 closed on the way. `pack` no longer takes a
`(track, frame)` predicate whose call sites each re-derived a feature from
an index — every track names the feature it follows, which deleted five
hand-maintained mappings. And `:dev-http` is gone: Django serves the page,
shadow-cljs only builds into the staticfiles tree.
227 CLJS tests, 31 Django tests, green.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Step 8's data model, ahead of its controls. Nothing here is a UI.
domain/params holds every knob's definition once — default, applicable area,
value constraints and the areas a change would force to regenerate. flow/take's
literal knob map becomes a view of it, so the take's defaults and the future
parameter panel cannot drift apart.
domain/feature adds subjects, features and groups as document data the renderer
never reads. A feature ID is stable for the whole clip, across occlusion: a run
of visible frames is not a new identity. An eye pair is an explicit group of one
or two eyes of the same subject, so a profile view with one identified eye needs
no invented partner. Settings resolve area -> subject -> group -> feature, and
dropping an eye from a pair materialises its effective values first so playback
does not jump. scene/problems now validates all of it.
Presence becomes per-feature rather than per-subject. freeze's :absent predicate
takes a track as well as a frame, so one occluded eye can be absent while its
partner still has a value; a full-face miss still marks everything absent. A
manifest may annotate known gaps as one-based inclusive intervals, which ingest
expands into observation tracks before measurement. An unobserved eye then gets
no vote in the iris pairing and cannot steer the shared gaze — gaze falls back to
whichever eye is visible. Temporal filters still see a sample on every frame,
held from the last observed one, because the numbers are a rectangular buffer;
the state mask, not the buffer, is what says the frame has no value.
js/app.js gets the same occlusion lesson: leading nulls from a face that starts
occluded used to throw away the whole take, and the neutral frame could be chosen
from a held duplicate pose.
Parameter editing, scoped regeneration and a feature-level detector remain. Until
one exists, footage without annotations falls back to the full-face mask rather
than claiming occlusions it cannot see.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B87NVmiU36qQmN9gmFYnJ9
`stabilize` is three things wearing one name, and it is now three functions in two
stages: `flow/measure/anchor` fits the rigid transform, `flow/condition` smooths
its parameters, `flow/measure/mouth` measures the lip rings through the result.
Parity is on the COMPOSITION and not on the pieces -- a split that agreed
function by function and not end to end would be a split rather than a port.
The oracle now drives `stabilize` at three configurations and the port agrees to
1e-9 on ref, rigid, transforms, outer, inner and aperture, plus `smoothContours`
at three radii. Two of the three configurations are at aspect 0.5625, a 1080x1920
phone clip, because at aspect 1 `pick` is the identity: a port that dropped the
anisotropy correction outright would pass every other assertion in the suite.
148 tests, up from 134.
Three decisions worth the reading time.
`makeXform` is not ported, and its absence takes the face oval with it. It
centres on the oval's bounding box and zooms until the face is 80% of the raster
height, so every vertex it touched carried a cropping decision made once, at
analysis time, from one frame's landmarks. Geometry belongs in the node's own
local space with the framing as a transform on a node, so this is a deletion. The
oval's only other consumer was the placeholder plate outline, which is painting.
The residual is taken against the RAW fit, and the prototype took it against the
smoothed one. That is the only deliberate numeric divergence here, and parity is
kept by asserting `anchor/residuals` on exactly what the prototype handed it. The
number's job is to say whether a section is stabilisable at all; folding the
smoothing error into it makes a slider look like a property of the footage, and
docs/architecture.md lists the residual under stage 3, which requires it to be
knob-free. `condition/anchor` therefore replaces `:transforms` and leaves
`:residual` alone.
The stage order is not the strict chain the table in docs/architecture.md looks
like, and that document now says so. The fit is knob-free, conditioning smooths
it, and the rings are measured *through* the conditioned transform -- so
`anchor avg` does re-run the ring mapping, which is a few hundred frames of twenty
points. The guarantee was only ever about the part that reads a source pixel, and
that part never sees a transform.
Two things fall out and are asserted rather than assumed. Smoothing and
subsampling commute, because both are per-slot, which is what lets `vertices`
stay a stage-5 knob downstream of a stage-4 one -- and it is also why the port can
smooth the full twenty slots where the prototype smooths eight and still match.
And `condition/contours` is `geom/moving-average` per vertex per axis rather than
its own clamped window, so "radius 2" cannot come to mean two different things at
the two knobs.
One dead end recorded so nobody walks it twice: the synth's head is perfectly
rigid -- its jitter is a whole-head translation, which a similarity absorbs
exactly -- so every frame's rigid configuration is congruent with frame zero's and
the Procrustes mean IS frame zero to 1e-15, jitter or none. "The reference is the
mean and not frame zero" cannot be asserted on this track and is asserted in
geom-test, where the two can differ.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Steps 2 and 3 land together because the model revisions in the middle changed
code from both, and splitting them now would invent intermediate states that
never built.
domain/channel value-at across framed/keyed/dense, plus a cursor
domain/node decomposed transform, composition order, time maps
domain/scene topological order, z paths, eval-frame and resolver
clock audio-clocked frame derivation, outside app-db
db/events/subs re-frame arrives; the playhead is document state
ui/player the rAF loop; reads, blits, dispatches (almost) nothing
ui/shell transport
133 tests, 1158 assertions. The scene plays at 30fps against audio, scrubs, and
runs at 1/4x through 4x; verified by driving a real browser over CDP rather than
by assertion.
Two evaluators, on purpose. `eval-frame` is the specification -- allocating,
order-free, obviously correct. `resolver` is what playback uses: cached topo
order and z paths, a cursor per channel, a preallocated point buffer per node.
Both run the same walk, parameterised only by how a channel is read and where
points are written, because two independent implementations of frame evaluation
would drift and the drift would read as a rendering bug rather than as two
functions disagreeing. scene-test asserts they agree frame for frame in forward,
backward and random order.
Deviations and decisions, each with a reason:
- raster/fill-poly! is now a thin wrapper over fill-poly-buf!, which takes a flat
preallocated buffer. ONE scanline fill serves the analysis stages, which speak
{:x :y}, and frame evaluation, which hands over a buffer it owns. The parity
suite still passes pixel-for-pixel, which is what makes the rewrite safe.
- The state mask carries ABSENCE ONLY. An earlier draft gave it a hidden bit too,
per architecture.md's "hidden flag + palette index", and that bit was a dense
[:vis] wearing a different hat -- two mechanisms for one question, which is how
a part ends up hidden by one and shown by the other.
- The palette is a parameter of evaluation, not a global. A node names a TONE;
which ramp that tone is read in belongs to the timeline it sits in.
- :over layers and a symbol :rate THROW rather than being ignored. Neither is
built and nothing can produce one, so this can only fire on data that has run
ahead of the code. A silently dropped override is a hand correction the user
made once, watched fail, and has no reason to trust again.
Three findings the model produced rather than received:
- Presence propagates asymmetrically. An absent transform drops the subtree; an
absent [:geom :pts] drops only that node, because an absent mouth outline has
nothing to draw but the head it hangs off has not moved. That asymmetry is the
reason presence is tracked per channel and not per node.
- Z paths need lexicographic compare, not `compare`, which orders vectors by
count first -- so a cel three levels under "a1" would jump in front of a bare
"a2" and the layer order would mostly work.
- A node stencilled by something that drew nothing is dropped, not drawn
unclipped: an iris floating over the cheek is worse than a missing iris.
docs/ revised alongside, and those revisions are the load-bearing part:
- A scene, a timeline and a symbol are one type. The doc had two structures with
the same fields and never said so. Two axes of nesting are now separated --
parent/child within a timeline is flat with parent pointers, instance nesting
is by reference -- which is why "nestable" and "flat" only sounded
contradictory.
- Palettes are named, live on the project, and are ENABLED on a timeline as a
channel. Absent inherits; present travels with the timeline, so a symbol
authored against :night stays night wherever it is placed. The output index
space is the concatenation of the named ramps, which keeps one buffer and one
flat table and incidentally stops two nodes in different palettes colliding on
a stencil.
- Stabilisation is a channel, not a mode: {s, theta, tx, ty} IS [:xform :*], so
the normalise on/off/per-plate toggle is which of the three channel shapes the
:head node carries. Always measure and always store factored -- smoothing and
velocity-minimum key selection both need the split to exist in storage.
- There is no camera node and none is needed. Placement is a node transform, the
stage clips what hangs off it, and project dimensions are independent of the
footage. `makeXform` is therefore not to be ported: it bakes a cropping
decision into every stored vertex.
- Export is removed. The .take writer was for an Animator Pro render script; the
target is encoding video in the browser, and step 9 now says not to port the
old one.
demo/swarm is 120 shapes on six orbits, entirely dense blocks behind store
handles -- the shape freeze produces at step 5, and the first thing to exercise
that path under load. It plays at 30fps, and bench-test keeps a deliberately
loose floor under it because a performance regression here does not announce
itself: the picture stays correct and merely arrives late.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PDfHGdV39zu6rvgbBTfDaT
Scaffolds frontend/ (shadow-cljs, reagent 1.2.0, re-frame 1.4.3) and ports
everything below the data model, with the JS kept as a numeric oracle.
domain/landmarks index tables, verbatim
domain/ring subsample, offset, simplicity
domain/geom similarity fit, procrustes, moving average
domain/raster indexed scanline fill, stencil, disc, rect
domain/palette the ramp, and the no-sampled-RGB rule
58 tests, 166 assertions. Parity with js/ on the identical 72-frame synthetic
track: fit-similarity, procrustes-mean, fit-residual, moving-average,
smooth-transforms, offset-ring and subsample-slots to 1e-9; the raster
pixel-for-pixel over the whole buffer.
Three deviations from the JS, each for a reason:
- synth.cljs jitters from a SEEDED generator, not Math.random. Parity is only
checkable if both sides can be handed the same track, and a failing assertion
has to be reproducible. `:rand-fn` takes the generator over, so oracle.mjs
stubs js/Math.random and js/ itself stays untouched.
- raster/->rgba replaces toImageData. ImageData is a DOM type and domain/ may
not touch the DOM; returning plain bytes also lets the
no-intermediate-colours assertion run in node. ui/canvas wraps it later.
- offset-ring lives in domain/ring, not domain/geom, per architecture.md: it is
an operation on an ordered traversal, not on a transform.
Step 1's "done" also names the swapped-iris vote, but pairIrises is in
pipeline.js and belongs to step 7. The precondition is asserted instead --
`:swap-iris` really does move both blocks -- so the vote will have a track that
disagrees with it when it arrives.
One finding, recorded in full in the test that measures it: smooth-transforms
buys nothing on the synthetic track. Against jitter-free ground truth, radius 1
helps by 17% on one noise realisation and hurts by 0.5% on another, so its
benefit is within noise; from radius 2 up the cost is unambiguous, and by radius
5 the filter is below the true motion's own high-frequency energy, i.e.
smoothing away performance. The test pins the shape of the knob rather than a
preferred value. This may say more about the synth's jitter being unrealistically
small (+/-0.001 normalised) than about the knob; step 6 settles it on real
footage.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A brow at 320x200 is fourteen pixels wide and three tall. Its shape carries
almost nothing at that size; its height above the eye carries the expression,
and a brow raise is the most legible beat on a face. So the ring is traced and
the height is quantised - the split the eyes already got, where the lid is a
traced feature and the iris a quantised primitive.
The decomposition is the point. The traced ring already contains the real
height, so adding a quantised raise on top would move the brow twice. The
height is measured OUT of the ring, quantised, and put back, so the shape that
renders is his at a height that snaps between a few levels and holds.
Measured at both ends rather than as one number, because raise and tilt are
different expressions out of one mechanism: both ends up is surprise, inner up
alone is worry, inner down is anger. They share a dwell - the gaze quantiser,
renamed quantizeSnap now that it has two callers - so the brow hits its pose in
one frame instead of crawling into it with one end arriving before the other.
Measured against the eye's corner midpoint, never its lid. Same trap the gaze
origin has and worth avoiding twice: brows and lids move together constantly,
so a brow that jumped on every blink would read as a tic. Rest pose from the
take median rather than the neutral frame, for the reason gaze learned the hard
way - that frame is picked by minimum mouth aperture and says nothing about the
brows.
Two correspondences resolved from geometry, not declared: which ring is which
brow, and which end is the outer one. The second matters more - backwards, the
tilt mirrors and worry renders as its own opposite, which reads as a directed
performance choice rather than a bug and would never be questioned. Which EDGE
is upper is deliberately left unresolved: it traverses the same ring the other
way, an even-odd fill has no winding, and both ends still land on fixed slots.
Also fixes a bug from the exposure work: the live render applied exposure to
the plate and the mouth but not to the eyes, so on 2s the preview and the
export disagreed. A preview that disagrees with the export is the one bug this
tool cannot afford. perfIndex now exists as a named thing so the two paths
cannot drift apart again.
91 -> 105 assertions. Ground truth on all four synthetic brow poses, tilt
separating worry from anger by sign, a blink not faking a raise, and a shared
dwell never emitting a half-raised brow.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three parts per eye, stacked the way the mouth is - dark lash ring, sclera
inside it, iris inside that, square pupil in the iris. A blink then costs
nothing: when the lid shuts the traced ring goes flat and the lash line
collapses to a lens, which is a closed eye, drawn correctly, for free.
Lids are a FEATURE, rotoscoped like the mouth: head-local, a key on every
frame, the same contour avg knob. The iris is a PRIMITIVE - a disc at a
quantised position - and that is where the stylisation lives.
Line of sight. Gaze is the iris centre relative to the midpoint of the eye's
two corners, in units of corner distance. Both corners are in RIGID, so the
origin and the scale are immune to the performance being measured; against the
lid ring's centroid instead, every blink would drag the origin down and fake a
glance at the floor on exactly the frames where the eye is most visible. Both
eyes share one gaze - at this size the difference between the two measurements
is noise, not vergence, and independent per-eye noise reads as wall-eyed
immediately. Openness stays per-eye so a wink survives.
Gaze is then quantised to a pixel grid with a dwell, which is not a
stylisation imposed on the truth: real eyes move in saccades, and the smooth
drift left in the measurement is tracker noise plus head-compensation error.
Snapping to a grid removes the noise and recovers the saccade in one operation.
The iris is placed in the frame of the already-smoothed, already-subsampled lid
ring - slots 0 and 8 of a 16-slot ring are the corners, and subsampling to any
even budget keeps them at 0 and n/2 - so it cannot drift relative to its own
eye. Size is authored from the take mean, never remeasured per frame: a radius
that breathes by a fraction of a pixel flickers a pixel on and off around the
whole silhouette. iris anchor toggles steady/free/locked, because how much the
eye wanders turns out to be an aesthetic choice and not only a correctness one.
Blinking gets hysteresis and a dwell like the teeth, plus one knob they do not
have: blink hold. A blink is one frame at 12fps and a single frame of closed
eye reads as a dropped frame, so once the eye shuts it stays shut long enough
to be legible. Detection accuracy is not the problem; legibility is.
The pupil is a square because at three pixels a circle is a plus sign with the
corners gnawed off, and it changes shape as it moves. Drawn from a rounded
centre shared with the iris so it is exactly its nominal size on every frame.
Iris/pupil clip by colour key against the indexed buffer, the way Animator Pro
would: the lid crops the iris at extreme gaze for free, so nothing has to clamp
the gaze, which would flatten the performance at the extremes that carry it.
Which iris block belongs to which eye is RESOLVED from geometry, not declared.
A swap looks almost right - each eye still has a disc roughly where it belongs
- so it survives an eyeball and then reads as a subtly wall-eyed character
forever. Voted across every frame; the test feeds a deliberately swapped track.
Also: exposure. Aesthetic sparseness was set by the extraction rate, which made
the timing a property of a directory of PNGs - auditioning 12 against 24 meant
re-ripping and re-detecting the whole clip. It is now a render-time grid, on
1s/2s/3s/4s, so the dense track keeps everything and the audio clock is
untouched. The take format already carried an exposure field; it was never
driven. Everything rides the same grid, because a head cutting on the odd
frames while the mouth cuts on the even ones reads as two performances laid
over each other.
41 -> 91 assertions. The load-bearing new ones: the iris pairing follows a
swapped track, a blink does not fake a change of gaze, a stencilled disc cannot
spill past its clip, a 3px pupil is 3x3 at every sub-pixel centre, and exposure
never reads a pose from the future.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The test renderer turned out to be the product. Everything that decides how the
work looks - stabilisation, reduction, timing, frame removal, palette - already
happens here, and the flat indexed output already reads the way it should.
The reason to leave is in the original design's own rule: never make a timing
decision that requires a full render to evaluate. Honouring that moved every
judgement out of Animator Pro, which left the host doing nothing but writing a
file, in exchange for modal UI, minutes-long renders, one-level undo, FLX delta
invariants, a single tween state and a cel singleton.
What does NOT change is the constraint. 320x200, indexed palette, flat fills,
no antialiasing - inherited, but load-bearing rather than accidental. The
rasteriser writes palette indices and expands to RGBA only at the end precisely
so nothing can soften an edge. Modern conveniences belong in the workflow.
Adds docs/design.md: the principles, carried over without the Poco/FLX/cel
machinery, plus architecture and an honest list of what is missing - the
largest gap being that plates still have nowhere to be drawn.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
