Repository navigation
Caustics and Temporal Resonator 0.1.1: the optical flow they never read is left out - #28
Merged
Merged
Conversation
…ad is left out Caustics and Temporal Resonator read only the luma and the clock from lib optical_flow(), yet each carried its Lucas-Kanade passes too. Those passes never cooked, but they sat in the network for anyone looking inside (board #4797). - lib: optical_flow(flow=False) leaves out the flow (lk8, lk4) and lum8p, which only lk8 reads. Caustics 0.1.1 and Temporal Resonator 0.1.1 pass it. BUILDING.md names the option. - Proof that nothing else changed: dev/probes/structure_diff.py against the builds before finds Caustics 7 operators fewer (lk8, lk4, their docked DATs, lum8p), Temporal Resonator 4 fewer (lk8, its docked DATs, lum8p) and only version strings besides. The other 23 effects are unchanged, so they are rebuilt for the lib hash with no bump. - The passes never cooked: the new dev/probes/flow_cooks.py pulled each 0.1.0 output with the clip, its depth and the stress beat for 120 frames, and lk8, lk4 and lum8p cooked 0 times. So performance and flashing are not re-measured. - Idle 0 everywhere: Caustics 0 of 227, Temporal Resonator 0 of 229, Showreel 0.1.26 0 of 7292, unwired and wired. The self-test and both repo tests pass.
kevinraymond
added a commit
that referenced
this pull request
Oct 29, 2026
… over it, its depth map the shapes they light and shadow, as real-time renderers light a scene (#29) Relight lights the live picture again with colored lamps that sweep over it, its shape from a depth map on input 3 (board #4677). - The shape: the depth as a height field at half size, Relief high, rounded inward by Roundness (a triangle average) and held to the depth at its edges, so a figure is a cliff over its backdrop. A slope of a few pixels down to a wall meters behind shadowed itself into a black outline. Normals take each slope toward the neighbor nearer in height, so the backdrop beside a figure keeps its own sheen. Without input 3 the picture's brighter parts bulge against their broad average, never sinking (that dug a moat). - The light: four looks (Stage, Studio, Sun with a blue sky, Two Tone), lamps on slow Lissajous paths hung above the relief's top, Lambert diffuse with falloff, GGX highlights, shadows marched over the height field at a quarter size, ambient occlusion after Luft et al. with a range check (without it every figure wore a black halo), light scattered under the surface (one Gaussian a color, red widest, after Jimenez et al.) and glowing through thin parts. - The picture's own light: Flatten divides its bright parts toward its average by the lower of its blurred and its own brightness, and never lifts the darks (lifting turned smoke and compression blocks into contours). - Music: each beat flares the lights through the flash gate, fading over 0.6 s (over 0.35 s, Up to 3 per second measured 3.5); loudness brightens them. - Flashing: Off adds nothing to any clip's own (seven clips, every look, the corners), Calm at most a flash a second, Up to 3 per second reaches WCAG's 3 on the dancer at 140 BPM and no further; red 0 everywhere through the red guard. - 4.6 ms GPU (95th percentile 5.6) and 0.5 ms CPU a frame at 1280 x 720 on the M2 Pro, the shadows' march the largest part; nothing cooks while unused (0 of 259). - Patents: clear by design, three changes made after the search. Canon's US 11,948,282 and 11,106,938 claim shading a region found in the picture by its normals: nothing is found or set apart, so skin isn't found by color and light scatters and glows everywhere. Adobe's US 8,379,021 claims shadows by the point of maximum elevation: a shadow multiplies its samples' light let through. NVIDIA's US 8,064,726 claims rendering with a sum of Gaussians: the broad smoothings are triangle filters (private/patents/relight.md; board #4805 asks whether other effects need checking against the last one). Also: Showreel 0.1.27 with twenty-six effects (idle 0 of 7553, Cycle 3.9 ms GPU, flash within each setting), the README row, status (Caustics and Temporal Resonator listed as 0.1.1, which PR #28 missed) and hero (2, eight rows of three, two), and BUILDING.md gains two TouchDesigner findings: a TOP's GPU time can carry another operator's, and the Blur TOP's filter types.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Caustics and Temporal Resonator read only the luma and the clock from
lib optical_flow(), yet each carried its Lucas-Kanade passes too. Those
passes never cooked, but they sat in the network for anyone looking
inside (board #4797).
lum8p, which only lk8 reads. Caustics 0.1.1 and Temporal Resonator
0.1.1 pass it. BUILDING.md names the option.
against the builds before finds Caustics 7 operators fewer (lk8, lk4,
their docked DATs, lum8p), Temporal Resonator 4 fewer (lk8, its
docked DATs, lum8p) and only version strings besides. The other 23
effects are unchanged, so they are rebuilt for the lib hash with no
bump.
each 0.1.0 output with the clip, its depth and the stress beat for
120 frames, and lk8, lk4 and lum8p cooked 0 times. So performance and
flashing are not re-measured.
Showreel 0.1.26 0 of 7292, unwired and wired. The self-test and both
repo tests pass.