Design revision refining the machine-altered-perception design (2026-06-05):
- Right axis is a LOCAL offline restyle (SD img2img on MPS), not a cloud API
(refines §4.1/§4.3, §9). Per-frame restyle boils/flickers — disqualifying for
a peaceful piece — so temporal coherence is now a hard constraint, met by
optical-flow keyframe propagation (the EbSynth principle, OpenCV impl).
- Content = a small NEUTRAL "scales of nature" library (~4-6 clips, one per
scale), not a single stitched cosmic-zoom film; keeps the neutral-base thesis
and interactivity while retaining the awe-of-scale richness (refines §6/§8).
- NEW: the scales form a navigable closed RING joined by short AI zoom/warp
transitions, driven by an infinitely-turnable endless rotary encoder; diving
past the microscopic wraps around to the cosmos (refines §2 selector / §11).
- i18n sharpened (§1.2, sharpens §10): Left labels are a Pi-rendered runtime
graphics overlay (annotation-track + per-language string tables, shaped via
Pango/HarfBuzz + Noto) — NOT a baked per-language overlay video, which would
break i18n economics. Added Pi-resolution headroom to open questions.
- Strict-PD sourcing map (NASA/NOAA/NPS/USGS); non-US terrestrial is the CC-BY
soft spot. Economics: local authoring is ~free, not $300-3k of cloud API.
Grounded in a local POC this session (M4 Pro, sd-turbo, OpenCV flow): all four
axes validated on real footage; deterministic axes ~2.4s, Right flow-propagated
~2.7min/8s clip and calm.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Ben Stull