Saltline is a good game and a better provocation. The number in the headline is real, but the interesting part is the zero: no images, no models, no audio files. The content is generated at runtime by code the model wrote.

That changes the production question. If the expensive part is no longer the assets, but the generator, and the generator is code, then a model that can write code can enter the pipeline. The bottleneck moves to judgement: does the result look right, does it collide, does it have a loop, does it feel finished?

My first attempt with the skill was Ironsight. It passed every gate it had — six of them — and still felt like a demo. That was the useful failure. The gates were measuring the frame and the loop, not the product. The second generation, Black Ice, made the product a gate and became a 318 KB game with four roles, a boss, settings, save, touch and a named death.

4generations of the shooter
8gates in the current harness
1question: can it be checked?

What this is not

It is not a template for Saltline, an engine, an asset pack, or a claim that a low-cost model has the same taste as a stronger one. It is a procedure and a measuring instrument. The result still depends on the subject, the art direction and the human eye at the end.

It is also not a benchmark. The A/B runs hold the model constant and change the harness. They isolate the harness, not the model. A different model, a different budget or a different subject will produce a different result.

How it was made

The project is a sequence of runs, each one changing one thing. The full evidence is in the cases; the short version:

Run 01

Sundown Circuit

City kart, no skill. A complete arcade racer at 890 KB. The control.

Complete, thin shell
Run 02

Sunset District

City kart, skill + gates. 1.58 MB, four laps, five rivals, settings, save, touch and a wet-road reflection system.

Use it
Run 03

Neon Breach

Shooter, no skill. 304 KB, five waves, four roles, a boss, grenades and a shield. Stylized, content-rich.

Content control
Run 04

Ironsight

Shooter, skill v1. 1.50 MB, 6/6 gates, 0.87 energy — one role, one gun, no shell. The run that produced the product gate.

Not enough
Run 05

Black Ice

Shooter, skill v2. 318 KB, 8/8 gates, four roles, a boss, two tools and a full shell.

Use it, with a budget

Credits and licence

Code, harness and prose are MIT. The game builds and screenshots remain © Norwin Yu; they are included as evidence, not as a grant of rights to the games. Saltline is the work of its creator and is referenced for comparison; its screenshot is a third-party capture. Three.js is MIT and is inlined in the city-kart builds with its licence notice intact.

The site has no analytics, cookies or third-party requests. Settings and best runs use localStorage on the player's machine. The games are single-file and make no network requests while playing.

cite this
@software{yu2026photoreal,
  author  = {Yu, Norwin},
  title   = {photoreal-game: a harness for photorealistic 3D games from code},
  year    = {2026},
  url     = {https://github.com/NorwinYu/photoreal-game},
  license = {MIT}
}

Build one

The most useful contribution is a case study with the failures left in.