Frames rendered on the lab RTX 5080 from the current code, as the four-wall room sees them. The wide strip shows all four walls side by side (left, front, right, back); the others show the front wall. Times are the mission clock: T-0:00 is liftoff.
Answered 31 Oct: 1B (checkpoints), size is no constraint, then tower look B (warmer amber), tower method A (refit on the real paint) and backdrop A (round 1 fixed). Running: BAKE10 and MATTE6. Proof of what each skill step contributed: Scene Loop Trial.
Recommended: B. The look is your call; the rounds in between are mostly measurement.
Recommended: A. The painted backdrop would bring the launch to about 105 MiB. The show runs from one PC's local disk, so that adds about a second of loading and costs nothing during the show.
Round 1's single painting looks best by far: wet concrete, reflections, lit equipment. Rounds 2–4 tried to fix its three faults (soft detail, a seam on the back wall, switching off one second before liftoff) by projecting AI tiles onto the ground, and each round looked worse: tiles drift, so half were rejected, and the rest leave blocks.
Recommended: A now. It is the version that already looks right, and its faults are specific. B can follow if the liftoff shows the painting is too flat as the capsule rises.
Both are AI repaints of our own grey render, so they keep its exact layout: steel stays neutral grey, the sky fill is cool blue-grey, and each lamp throws a warm pool on its deck and stairs, with dark bays between. A has tighter, cooler pools. B is warmer on the inner faces and closer to the concept art.
A new method is running: instead of hand-tuning lights, which failed twice, it solves for the light settings that best reproduce the option you pick, and checks the result from a second window it was not fitted on. It fits both options.
Recommended: B. It is closer to the concept and to the warm look you liked at ignition. The run currently bakes B.
The baked light maps are now installed (shrunk from 25 MB to 6 MB) and the tower was relit toward the concept art. Below: the tower in the left window at T-1:15, the concept, the version before baking, and now. Measured against the concept: brightness 0.030 → 0.066 (concept 0.083), warmth +0.10 → +0.26 (concept +0.28), area crushed to black 17% → 6% (concept 3%). Frame times stay within budget on the four walls (≤ 4.2 ms); the desk room is unchanged at 4.3–4.9 ms around liftoff. All 1,390 tests pass.
What is still wrong: the warmth sits in the steel, so the whole lattice reads peach. In the concept the steel is grey-white, and the warmth comes from the lamps, their pools of light and the yellow rails, with darker gaps between the pools. The next round keeps the steel neutral, puts warm light only where the lamps are, and makes the lamp heads glow.
Out of the dark, warm, lamps and rails visible. Too even, and the steel is the wrong colour.
The flame light now turns the tower warm from below, as in the concept.
The painting is now 54 AI-repainted tiles at screen resolution, projected onto the real ground shapes so it holds through liftoff, with a second, flame-lit version for ignition. The automatic score was fixed and now agrees with my judgement on all 14 test pairs. Not in the show yet: several tiles have visibly different brightness (the rectangles in the sky below), it is 61 MB against a 60 MB budget, and the worker's room renders put the painting in the wrong place, so there are no usable room frames yet. The next round evens the tiles, fits the budget and renders every view.
Wet concrete, lit roads and the coast are there; the tonal blocks along the top are the seams to fix.
The correction brings back the warm dawn glow and a level sea. It still changes some hardware, so the real geometry stays the authority.
We rendered our pad layout from the cabin, had AI repaint it as a realistic night scene, and put the painting behind the real 3D tower and steam. Below are RTX frames with it switched on. It also makes frames faster: the slowest liftoff check on the desk drops from 4.8 to 3.7 ms.
Not in the show yet. The painting is two to four times softer than the screens, there is a thin seam on the back wall, and it switches off one second before liftoff (last image), so the view cuts back to the grey slab. A follow-up worker is repainting at screen resolution and projecting the painting onto the real ground shapes so it holds through liftoff, with a flame-lit version for ignition. Its automatic score had judged this version worse than the old pad; that score was wrong, and it is being fixed and tested against my judgement before it is used again.
Reflections in the wet concrete, lit equipment and floodlight pools, where there was a flat grey slab.
Coastline, water and the radar dome in the distance; still soft at full size.
The follow-up carries the painting through liftoff.
The tower was lit in Blender with path tracing (the sky, the floodlights and all 171 work lights, bouncing between the steel members), and the result was saved into light maps the show reads at no extra cost. Below: the tower alone, left window, before and after, next to the full ray-traced render of the same view, which is what correct lighting looks like. Similarity to the ray-traced render rose from 0.43 to 0.62 (1.0 means identical).
Not in the show yet: the light maps take 25 MB more than the launch's download budget allows. A follow-up worker is shrinking them, installing them, and relighting the tower to match the concept art below, which is warmer and brighter than our ray-traced render: yellow floodlight pools, glowing rails, steel turning orange at ignition. That one is scored against the concept art too.
Light now falls off between the members and the rails pick up warm light. Still too even and too grey.
The flame light now fills the structure from below instead of painting stripes. It should be orange, not cream.
The next round lights the tower toward the left column: brighter, warmer, with pools of floodlight.
All round-3 work is merged: ignition light, tower work lights, pad dressing and the shoreline repair. All 1,365 tests pass. Speed: 53 of 61 liftoff checks pass; only the desk room at 85–93 s is still slightly over (4.5–4.9 ms against 4.5).
Better: at ignition the tower now glows warm and the frame gets warmer, not darker; the steam on the side walls is bright and full again; the twilight sky is right. Still wrong: the pad is a wide grey slab with a few small boxes, like a model railway; the steam seen through the front window is a flat tan smear with a ragged edge. Next to the AI targets it does not yet look real.
Tower work lights and floodlight heads show; grass banks frame the hardstand.
Sparse, small and clean. No wet concrete, no beams, no lit roads beyond.
The tower is lit warm by the flame and the frame no longer goes dark.
The steam in this view is a flat tan smear with a ragged edge.
Side-wall steam is bright, full and warm again.
Still lighter and grainier than the gold, glowing target below.
The AI concept the ignition worker drew from our actual camera view.
The tower lights and pad dressing as they first came in, before the ignition light was merged.
Left: AI target. Right: the pad-dressing result (software preview). The target is what a real night pad looks like from 100 m up: wet reflective concrete, rows of lit equipment, floodlight towers with beams, lightning towers, tanks, roads lit to the horizon. Ours is a grey plain with a few boxes.
All three fixes are merged: the launch bugs, the pad surroundings and the tower finish. Every bug from this morning is fixed: steam stays outside the windows, the floating rectangle is now a floodlight, the T-8 blob is gone, the tower no longer blows out, the dotted clouds and the sawtooth coast are gone. Speed improved a lot: the four-wall room now passes the liftoff budget everywhere; the desk room is just over (4.5–4.9 ms against 4.5).
But it still does not look real, and the launch itself got weaker. The pad is now a huge empty grey slab, and ignition and liftoff went dim and grey: no flame glow, no golden steam. Compare it with the AI target at the bottom of this section.
Sky and frame are clean; the rectangle is gone.
The pad is a vast, empty grey plain lit by a few white smears, like a car park in snow. The floodlight in the foreground reads as a sign on a stick. Nothing like the dense, warm-lit pad in the concept.
The pre-dawn sky is now right: a smooth deep blue to lavender to warm glow, with soft clouds that no longer pop in. The steam blob is gone.
The pad is still empty, and the thin vapour the fix added is too faint to see from the front window.
The tower is readable steel, and nothing draws over the frame.
The whole view goes dark. There is no flame light on the pad or the tower, so the most dramatic moment of the show looks like a power cut.
Steam stays inside the windows on every wall, and the tower is no longer a white slab.
The steam is now dim grey-brown instead of glowing; it lost the light that made it the best part of this morning's frames.
The sawtooth coast and the camouflage tiles are gone.
The hardstand reads as a giant octagonal platform with a black rim on a flat brown plain: clean, but not a place.
A clean beach line, sea and dawn sky; this frame is close to real.
Shallow water near the shore is blocky; the land is still soft.
The dotted curtain streaks and the orange smudge are gone.
The clouds are now smeared flat streaks. A worker is fixing an abrupt cloud change at 1:54 found in review.
The AI concept the launch-bugs worker started from (ignore the orange rocket it invented: from the capsule you do not see the core stage). This is the bar: warm work lights everywhere, floodlit vapour, gold steam and flame light at ignition.
Rendered on the RTX 5080 with the new tower, ignition, lighting and windows merged. Honest read: the opening does not look real. Some defects are bugs that make it look broken; the rest is a pad environment that looks like an old video game. I checked each one against the older code: almost all of them were already there, so today's merges did not cause them, but they did not fix them either. Three workers started on the fixes this morning.
The window frame is clean. The tower has real structure (stairs, grating, braces).
Wrong: the sky is a flat orange strip with a hard top edge under a black sky. The pad is an empty grid with lamp pools, like a car park. A white glowing rectangle floats at lower right (a bug). The tower is too clean and evenly lit. From here to T-0:30 almost nothing changes.
Wrong: an opaque, skin-coloured steam blob with a hard edge, and the ground under it lit with a straight cut. The blob is new today (the ignition package starts the steam earlier). A full cloud deck appears that was not there at T-0:30, and the clouds are dotted noise.
Wrong: the steam is drawn over the window frame and the clock (bottom), as if it were inside the cabin. The tower turns into a glowing white slab. The pad is a blotchy camouflage of square tiles.
Left, front, right and back walls. The steam billows on the side walls have real volume and scale; they are the best part of the scene.
Wrong: the steam spills over the frames, the tower is blown out, and the back wall is a smeared ground.
Wrong: the shoreline is a row of sawtooth spikes, the land is a smeared photo with dark holes, and the pad tiles still read as camouflage.
The sea and the dawn glow on the horizon are close to right.
Wrong: the coast sawtooth, and dotted clouds overhead. At T+0:22 the clouds smear into dotted streaks with an orange smudge on the sea; that was in the older code too.
The first of the three fixes is in: worn paint, darker grating, railings, gusset plates, cables and level numbers. Tests pass and it costs no frame time. An independent review found no defects.
Honest read: a modest step. The steel reads as grey painted metal with real stairs, but it is still far from the concept. What makes the concept look real is light: dozens of warm work lights on every level, lit floor pools, and a glowing pad and coast behind it. That is the next tower round, after the pad-environment fix lands. The handrails also came out orange; they should be safety yellow.
Earth's curvature, the thin blue limb, the sun just risen and a cloud sea lit pink from below the horizon.
Known: blurred orange patches on the left; the clouds read slightly like popcorn up close.
Earth from low orbit with real NASA surface and cloud maps and a sun glint on the ocean.
Earth filling the window, NASA 8k clouds, Antarctica and South America, a soft terminator.
Real NASA photographs draped on the real terrain under a physically correct Sun: the central peak, its shadow, craters and texture at one sharpness across the floor.
The lunar horizon from real photographs and a half-lit Earth above it, as in the Apollo 8 Earthrise picture.
The cloud bands come from real spacecraft imagery and look rich, with Io crossing the disc.
Defects: Io's shadow is a hard black cut-out with no soft edge; a faint dotted texture seam runs down the left; the disc looks flat (no limb darkening).
The weakest part of the show. Near the ring plane the rings turn into a flat, opaque white sheet with a hard gap; ring particles are oversized balls; the planet shows through the rings as a strange translucent band.
Next: a dedicated Saturn round after liftoff and Jupiter.
The show still ends at Saturn. The homecoming journey (Saturn back to Earth) is designed but not built yet.