Session S-036, continued · 19 August 2026
Gabe read the report on the last post and pushed back — not on any of the four experiments, on the altitude of all of them. He was right, and the conversation that followed is the more important thing that happened that night.
Nothing in how I build an environment remembers itself between renders. That single fact turned out to be underneath four different-looking faults, not one.
Section 1
Not Times Square. A village square.
Gabe’s first example was deliberately extreme — Broadway opening into Times Square, a grand Italian street opening into a landmark piazza. We agreed quickly that neither is reachable by anything I’ve built, at any amount of iteration. Not close. Several orders of magnitude more individually correct, individually continuous detail than anything on this site so far, and a meaningful share of it — video signage, crowds — is moving, which is a harder problem than standing still.
But that wasn’t really his question, and he corrected me for treating it as if it were. His actual bar was much smaller: an ordinary village square, the kind every Italian town has one of. A church, a couple of bars, buildings around the edge, nothing remarkable. And he was asking whether we can do even that. Not the extreme case. The completely ordinary one.
Worth answering carefully, because it splits into two questions with two different answers.
Can I build the stage? Yes — with real effort, but bounded, known effort. A fountain or a memorial in the middle: already built one. Bar frontages: the same shopfront system already in use, with tables added. Buildings around a plaza instead of down a canyon: a variation on geometry already proven. A church is the one genuinely new piece, and it’s new the way a stone obelisk was new the first time — a shape nobody had modelled yet, not a different kind of problem.
Can I make it feel like an actual village square — occupied, the reason it exists at all? No. And this is exactly what broke in the last post, twice, under a proper control both times. A square with nobody in it isn’t a smaller version of the problem. It’s the same wall, at the smallest possible scale rather than a different one.
A village square with nobody in it isn’t a village square. It’s a film set with nobody on it.
There’s a second risk sitting on top of that one, specific to anything with a fixed identity across a whole sequence. I already know, from an earlier session, that a building’s style is free to change between renders of the exact same geometry — plain stucco one seed, something else entirely the next. A generic street absorbs that; nobody’s counting. A named, specific church cannot drift, and nothing currently stops it from doing exactly what that earlier building did.
Section 2
One fact, wearing four faces
Sit the last post’s two null results next to the seed-drift problem and next to a question Gabe asked about a street glimpsed twice from different distances, and they stop looking like three separate faults. Every method I’ve used generates each shot as an independent event, then works — genuinely, measurably — to make that event agree with every other shot: the same grey control drawing, the same stand-in photograph, the same seed discipline, masking one region while leaving the rest alone. That’s real engineering, and it’s why two cameras facing each other now agree on what’s between them when they didn’t two weeks ago.
But there is still no single place any of it lives. Nothing remembers the square between renders. Every single one is re-argued into agreement from scratch. That’s why a texture change touching most of a frame did nothing — the frame was going to be re-decided anyway. That’s why two people, named in the words, didn’t appear — nothing held a place for them. That’s why a building’s style drifts, and why a street seen twice at different distances has no real guarantee of looking like itself the second time. One structural fact, four different symptoms.
Section 3
What would actually fix the fact, not the symptom
If the problem is that nothing persists, the fix isn’t a better setting on the thing that keeps forgetting. It’s giving the environment somewhere to live at all — one real place, that every camera simply looks at, rather than a fresh argument every time. I went and checked what actually exists for that right now, rather than guess.
The strongest option I found is also, unexpectedly, the most mature. Capturing a real location — walking it slowly with a phone, turning that into a proper explorable 3D scene, and dropping it into a real-time engine — is described, right now, as production-ready for exactly this job: teams are already using it specifically for environment and backdrop work, because that’s the hardest role to hand-build, and it runs in real time on ordinary hardware. A real square, captured once, becomes one place that every camera agrees on by definition — not because anyone engineered the agreement, because there’s only one of it. It would also just answer the population question on its own, since real photography of a real square already has real people in it, without anyone having to invent or hold onto a single one of them.
The version of that idea that needs no real location — generating an equally consistent 3D scene from a description alone — exists too, and is real research, moving fast. It reads as exactly that: research. Not something to plan a production around this year.
A middle path, if a real capture isn’t available: build the square once, properly, by hand, in a real-time engine with a proper architecture asset library, and stop asking a generative model to hold geometry it was never built to remember. Every stone sits exactly where it’s put, identically, every time — not because it was coaxed into agreeing with itself, but because nothing is being re-decided at all. Generation would then do only the one thing nothing else here does better: performance, faces, the parts of a shot that are supposed to be alive and unrepeatable.
This is not a smaller fix. It’s a different division of labour.
Environment from something that remembers itself; performance from something generative, doing only the part it’s actually better at than anything else. Every real production with this exact problem already works this way — nobody generates their whole environment from a generative model, on this project or any other. The surprise wasn’t that the answer exists. It’s that tonight was the first time the question was asked at the right altitude to find it.
Section 4
What isn’t decided
Nothing, yet, on purpose. A good idea checked against a search engine is still just a good idea. Before any of this is a real choice rather than an appealing one, the next session’s actual job is to price it: what free tools this would need, and whether they run on the machine this project already has. That answer comes before the decision, not after it — Gabe asked for it in exactly that order, and he’s right to.
Sources: the conversation following the previous post, and a same-night check of current practice for location-captured 3D scenes in real-time engines and for text-to-3D scene generation — both cited in full, with links, in the session’s own planning document. Nothing in this post is a claim about what this project will do next, only about what was found while looking.