Voyage Demo Progress

Demo targets

T1	Plays first time in the room (one wide video, launch to Moon, 0–399 s). Venue kit tested on a lab PC with the same 4080 as the venue.	Your A/B test
T2	Liftoff looks real, 0–125 s: background, tower, smoke, clouds (the four hard parts below).	Running
T3	Climb and orbit believable, 125–290 s: clouds, night continuity, space fixes.	Running
T4	Moon swing-by tells its story, 290–370 s, plus the time-lapse to 399 s.	Running

Rule (owner): no full-show render until each hard part passes in its own short clip. Look rule: a movie, not a replica; push what the eye reads, compensate for room light.

Compare: AI target vs ours

Same moment, same wall, same framing. Night frames are dark on purpose; use the buttons to brighten every picture in this section equally so the details can be inspected. Click any picture for full size.

Brightness:

Tower, 60 s, front wall (night, before ignition)

AI target (board 1 B)

<b>Ours now</b> (after TOWER3) — **Ours now** (after TOWER3)

Look at:

The tower on the left: in the target, bright pools of light and hard dark gaps on the steel; ours is evenly lit.
Warm lamps against cold floods: two clear colours in the target; ours is one neutral colour.

Gap: large. Targets set for TOWER4: pools 6:1 (bright vs dark parts of the tower). Next evidence: TOWER4 + real-GPU render, about 3 h.

Tower, 88 s, front wall (ignition)

AI target (board 2 B)

Look at:

The steel facing the flame: gold in the target; ours is pale cream.
The side away from the flame: deep shadow in the target; ours is almost as bright as the lit side.

Gap: large. TOWER4 targets: red at least 3× blue on lit steel; lit side 4× the shadow side.

Tower, 92 s, left wall (liftoff)

AI target (board 3 B)

<b>Ours now</b> (after TOWER3; old live smoke) — **Ours now** (after TOWER3; old live smoke)

Look at:

The tower on the right: strongly gold-lit with dark structure in the target; ours is white.
The smoke: big gold-lit billows in the target; ours is the old live smoke (film smoke not in yet).

Gap: large on both tower and smoke.

Background, 60 s, front wall (far land and sky)

AI target (board 1 B)

<b>Ours, best repaint so far</b> (round 2) — **Ours, best repaint so far** (round 2)

Look at:

The horizon band just above the far grass: the target has rows of warm lights and a warm glow; the repaint added a thin line of small lights, the original has a few.
The sky: shaped dark clouds in the target; the repaint flattened them (worse than before).
Use the "Brighter" buttons: the differences are in the dark band.

Gap: medium on lights, worse on clouds. Next: BG1d (keep clouds, add town glow, all three walls) + live lights (BG4b).

Background, 60 s, all three walls: round 3 repaint (BG1d)

AI target (board 1 B, front)

Round 3 front — **Ours, round 3**, front

Round 3 right — **Ours, round 3**, right

Look at:

The sky: cloud shapes are kept this time (round 2 flattened them).
The warm glow: now on all three walls, but it sits up in the clouds like a sunset streak, not on the horizon line where town lights would be.
A small stray mark at the far right of the front wall's horizon.

Gap: medium. Next: glow onto the horizon line, remove the mark; lights come from the live light layer. Not in the show yet: installing the painting into the backdrop is a separate step once it is right.

Background, 105 s, front wall (climbing at night)

AI target (board 5 B)

<b>Ours now</b> (default) — **Ours now** (default)

Look at:

The target shows the coast from above: dark ocean, a thin line of city lights, a warm band at the horizon.
Ours: almost black (default) or the pad painting wrongly drawn from low down (blending on).

Gap: very large; this is the worst moment. Fix workers BG3b (view) and BG2b (night ground with lights) running.

1. Background: launch site and beyond (0–130 s)

Owner rating: still view on the pad ≈ 70%, once the cabin climbs ≈ 30%. Goal: 70% everywhere from parts A–C, the last 30% from part D. Design: docs/voyage/design/background.md.

A. Far painting (backdrop)

A1	Focus and skip list for the pad background. Focus: flat dark Cape, clean horizon, many distant lights in rows with glow, reflections in the river, warm town glow, shaped night clouds. Skip: 3D beyond the apron, detail inside dark land, exact building shapes, stars. Pushed: more and brighter lights, stronger glow, deeper black.	Agreed
A2	Tool: open SDXL model on our lab GPUs, locked to our edges and depth, whole wall painted in one pass (no tile seams). Free licence. FLUX ruled out (non-commercial). Research and worker agree. The 3 weak checks the review found (horizon, seams, dim lights) are fixed (BG1b).	Done
A3	Test: repaint the pad's front wall at 60 s at full resolution. First run: technically clean (0 protected pixels changed, no seams, 8 GB GPU) but it did not paint: a slightly smoothed copy, lights fell from about 1,050 to 550, no town glow. Cause: very dark input, low repaint strength, tight geometry lock.	Did not paint
A3b	Second round, 17 variants on the lab. Best: SDXL at strongest repaint plus a separate light layer: a thin line of distant lights along the whole horizon, light rows and reflections. But the sky's cloud shapes were flattened and there is no warm town glow yet. Qwen made blocky garbage and is ruled out.	Done
A3c	Director decision: AI paints the base (land, water, sky with shaped clouds, warm town glow); lights come from the live light layer (part D) so they are not doubled. Next round (BG1d) keeps the clouds, adds the glow, then paints all three walls as one.	Running
A4	All three walls painted as one continuous view.	Next
A5	All-round night sky (SKY1): done and now the default; real-GPU clip 100–136 s queued.	Done (code)
A6	Climb horizon painted for a few heights.	Planned

Our pad view at 60 s, three walls — Now, 60 s, three walls: few lights, flat sky, smudged far land.

Agreed pad look target — Agreed look target (pad, board 1 B).

Front wall before repaint — Front wall at 60 s, before.

Best repaint so far — Best of round 2 (SDXL, strong repaint + light layer): light line along the horizon; clouds flattened.

Best repaint crops — Round 2 best, 100% crops: sky, horizon lights with reflections, far-right light rows.

Qwen crops — Qwen-Image-Edit, same crops: blocky garbage in the far band. Ruled out.

First repaint crop 1 — A3 first run, 100% crop of sky and horizon: input | SDXL dual | SDXL union. Almost no change; the small aircraft light is lost.

First repaint crop 2 — A3 first run, 100% crop of far land and water: no added lights or glow.

Repaint mask — Test A3: green is the only area the AI may repaint.

B. Painted textures on the 3D models (worker BG2)

B1	Inventory of 3D pieces and their sharpness (BG2).	Done
B2	Projection-painting tools built and tested on a fixture; the round-trip on the real show is being set up (BG2b).	Running
B3	Night map of the Cape from NASA Black Marble: the real coast and towns are right, but blocky (500 m squares) and dim. BG2b turns it into crisp street lights along the real roads.	Running
B4	Check: short real-GPU clip of the climb; no popping, seams or crawling.	Planned

Climb view at 105 s — Now, 105 s: dark daytime ground, no city lights, flat ocean, hard edges.

Night coast target — Target (board 5 B) without the two ground smoke columns.

C. Blending painting and 3D (worker BG3)

C1	One haze on paint, ground, water and cloud (the main blender). Built behind a switch.	Done (code)
C2	One night light on paint and 3D. Built; the painting's own baked lighting is a remaining limit.	Done (code)
C3	No hard join line: 2° feathers on the survey edge and the painted backdrop. Built; needs the real-GPU look.	Done (code)
C4	Room-light preview (8% black floor, −15% colour) and corner measuring tool. Built.	Done (code)
C5	Hand-over from painting to 3D between 120 and 720 m height; smooth (0.5% per frame), done before 110 s.	Done (code)
C6	Three walls match at the corners (≤ 3% brightness step).	Measuring
C7	Real-GPU comparison, blending off vs on. Result: at 105 s with blending on, the pad painting is still drawn from a low viewpoint with dark wedge shapes; with it off, the view is almost black. At 97 s both show dark wedges cut into the apron. Not ready to switch on; fix worker BG3b running.	Problem, fixing
C8	Independent review of blending: the ground gets hazed twice, and water loses its night glint. Goes into the next blending round after BG3b.	Problem

105 s blending off — 105 s, blending off: almost black (the old black-out).

105 s blending on — 105 s, blending on: pad painting still drawn from a low viewpoint, dark wedges.

97 s: dark wedge shapes cut into the apron (both settings).

D. Life: lights and motion (worker BG4)

D1	Distant lights: bright core plus glow, gentle out-of-step twinkle (≤ 9%), red beacons blinking once a second.	Done (code)
D2	Light streaks shimmering on water; floodlight reflections in the pad puddles.	Done (code)
D3	Slow water shimmer; existing pre-launch steam vents reused. Cloud drift is with the night-sky worker.	Done (code)
D4	Flicker added to the flame light on ground and backdrop; soft smoke shadow kept.	Done (code)
D5	Real-GPU comparison, lights off vs on: practically identical; only one beacon differs (blink phase). Built, but far too subtle to see. Worker BG4b is pushing every element until it clearly reads, also in the room view (target: at least 3× as many visible lights, visible reflections).	Too subtle, fixing
D6	Independent review of lights: new light sprites skip the shared haze when both switches are on; tests too weak. Goes into the next lights round after BG4b.	Problem

60 s lights off — 60 s left wall, lights off.

60 s lights on — 60 s left wall, lights on: no visible difference. Too subtle.

2. Tower: structure and light

Agreed focus: the big readable structure (legs, braces, platforms) with crisp edges; flame light warm on the near side, hard shadow behind, flickering with the flame; floodlights with glow. Skip: bolts, rivets, small pipes, cable trays, signs, exact lamp count. Design: docs/voyage/design/tower-look.md.

2.1	Movie never asks for missing tower files (a review found the movie would hang).	Done
2.2	Close-up steel sharp enough: needed ≥ 2 texture pixels per screen pixel, was 0.07; now 7–8.6 at the 60/86/92 s cameras.	Done (code)
2.3	Flame light follows the rising flame from 86 to 95 s (was frozen at one moment).	Done (code)
2.4	Lab re-bake installed (28 lighting maps) and checked on the real GPU at 30/60/88/92 s: structure passes (grey steel, crisp close detail, no longer reads as wood).	Done
2.5	Flame light: found why it was weak (the bake let through 1.5% of the flame, and the engine is at 18% power at 88 s). Now about 17× stronger at 88 s and 9× at 92 s, with flicker. Movie uses the new tower again.	Done (code)
2.5b	Night re-bake with aimed floods installed and checked on the real GPU: at 88/92 s the tower is clearly brighter where the flame hits, but pale cream-white, not gold; at 60 s still evenly lit, no visible floodlight pools. TOWER4 now has measured targets: gold (red/blue ≥ 3), lit/shadow ≥ 4:1, pools 6:1.	Better, not there
2.6	Your rating of the tower, static and at liftoff, and what bothers you most.	Needs you
2.7	Life on the tower: beacons, lamp glow, haze, shake at ignition (with background D).	Planned

Tower at 60 s — Before, 60 s front: even cream colour, reads as wood.

Tower at liftoff — Before, 90.4 s front.

New tower 60 s — After re-bake, 60 s front: grey steel, crisp detail; still evenly lit.

New tower 88 s left — After re-bake, 88 s left: detail good, no warm flame light yet.

New tower 92 s — After re-bake, 92 s front (old live smoke): tower still white, not gold.

Tower 88 s after TOWER3 — After TOWER3, 88 s left: brighter where the flame hits, but cream, not gold.

Tower 92 s after TOWER3 — After TOWER3, 92 s front (old live smoke).

Ignition target — Target (board 2 B): gold engine light on the steel, cream steam.

3. Smoke: liftoff (83–112 s)

Agreed focus: fills the lower window within seconds; white steam south, tan exhaust north; big billows with dark folds; gold from below, dark blue on top; always rolling. Skip: wisps, fine swirls, exact physics, smoke hidden behind the tower. Film-quality smoke is simulated and rendered offline, then put into the movie. Design: docs/voyage/design/launch-smoke-film.md.

3.1	Pipeline: all 7 lab GPUs set up; camera alignment with the show passes (S1).	Done
3.2	Lighting calibration (S2) against the new tower bake: queued on the lab together with a fresh camera alignment (S1).	Queued
3.3	Path A, GPU simulator: the 4 review problems are fixed (SMOKE-FILM2b). Its first real run waits for the light calibration (3.2).	Done (code)
3.4	Path B, whole near field rendered in Blender: tools built; waits for a simulation and S2.	Next
3.5	Path C, Mantaflow: pilot fine, but full run projected 25 h; parked.	Parked
3.6	Compare paths at 88/92/98/105 s against the target; you judge a short clip.	Planned

Smoke now — Now, 92 s front (old live smoke).

Smoke target — Target (board 3 B): big gold-lit billows, dark folds, deep blue night.

4. Clouds (108–135 s)

4.1	Cloud floor at 126 s has real shape (no repeating pocks). Checked on the real GPU.	Done
4.2	Inside the cloud at 116 s: still flat; the worker's method stopped improving (contrast 0.24 stops, needs 0.30). Needs a different approach. Real-GPU clip 110–128 s rendering to judge.	Problem
4.3	118 s: smeared surface and combed spray on side walls (CLOUD1d).	Running
4.4	Brightness jump entering the cloud at 114 s: 17.4% → 2.8%. Near/far hand-off error cut about 50×. Checks can no longer pass falsely (CLOUD1d).	Done (code)
4.5	Night above the clouds (SKY1): the cause was the old dawn Sun peeking over the horizon as the cabin climbed; now held well below until 130 s, no Sun disc, moonlight from above, gold engine light on nearby clouds. Default. Real-GPU clip 100–136 s queued.	Done (code)

Clouds at 126 s — Now, 126 s: good shape, wrong time of day.

Clouds at 118 s — Now, 118 s three walls: smeared surface, daylight.

118 s target — Target 118 s (board 6 B): gold near clouds, dark blue far.

126 s target — Target 126 s (board 7 A): the gold fades, dark cloud ocean.

5. Space (135–290 s)

5.1	Orange blob around the Sun removed; unverified thruster cones removed; ocean sheen above 10 km.	Done
5.2	Steering at 228 s was 29° off, moving the sunrise; now under a millionth of a degree, sunrise at its target height.	Done (code)
5.3	Sun-halo fix now limited to space beats; the climb and Moon skies keep their look.	Done (code)
5.4	Turn at separation was 5.5°/s; now 1.1°/s (comfort limit 4.8).	Done (code)
5.5	Real-GPU check at 125–300 s done: night Earth with city lights and a sunrise sliver at 240 s; full Earth at 285 s.	Done
5.6	Green airglow restaged (SPACE1f): now within 1–4 px of the Earth's edge on every wall (was 45–62 px).	Done (code)

Space 228 s — Now, 228 s front: night Earth, city lights; the green arc above the limb is the 5.6 problem.

Space 240 s — Now, 240 s front: first sunrise sliver at the centre of the horizon.

Space 285 s — Now, 285 s three walls: full Earth in daylight on the front wall.

6. Moon swing-by (290–370 s) and time-lapse (to 399 s)

6.1	Violent snap at 383 s in the time-lapse fixed: fastest turn now 13.7°/s (limit 14).	Done
6.2	Storyboard targets: the Earth now enters the left wall cleanly at 319 s. The other numbers were measured against the flight path; director decision: keep the path and the story moments, accept the measured values (Moon about 88° across at 344 s instead of 70°; Earth hidden earlier, on the back wall only, which Algoma does not have).	Decided
6.3	Sharper Moon terrain along the flown track (NASA LRO data, about 100 m per pixel) and a larger Earth at Earthrise (up to 2× true size so it reads on the projector) (MOON1b).	Running

Moon 320 s — Earlier draft, 320 s front.

Moon 350 s — Earlier draft, 350 s front.

7. Playback in the room (T1)

7.1	Venue kit tested on lab pc1 (same 4080 as the venue): the 8192-wide HEVC file plays 15.8× faster than real time with no dropped frames; the three separate wall files play together. The 7680-wide H.264 version fails GPU decoding and is not used.	Done
7.2	Your double-click test on pc1 or pc2 (A/B).	Needs you
7.3	Portable player as a backup.	Planned

8. Sound

8.1

Same rule as the picture: push what the ear reads (ignition rumble, crackle, wind through the climb). After the visuals pass.

Planned