NINE LOOPS: SUBWAY – an Exit 8-style anomaly loop game in three.js (mobile, EN/JA, online leaderboard)

Hi all! DESIGN OFFICE NINE here, a small web studio in Japan. This is NINE LOOPS: SUBWAY, a loop-escape game built with three.js.

:play_button: Play: NINE LOOPS: SUBWAY|NINE CITY
Free, no install, PC + mobile, English / Japanese.

The rules

You’re stuck in a subway tunnel that keeps repeating.
See an anomaly? Turn back. Nothing wrong? Keep going.
Get it right 9 times in a row to find OCHIYO, the girl in the red hood. Get it wrong and you start over.

It’s inspired by The Exit 8, but the anomalies here are mostly the residents of the tunnel. NINE CITY is a city made from discarded emotions, and its residents are echoes of animals that were hurt by people. Some of them are friendly, some aren’t, and telling them apart is the game.

How it’s built

  • Plain three.js + ES modules, no framework. One HTML file and a folder of GLB assets.
  • The tunnel is a fixed track of segments: a sign straight, S-curves, then a 42 m “anomaly straight”. Each lap swaps in a new encounter. Turn around and walk back far enough, and it counts as “turned back”.
  • While you walk the current lap, the next encounter and its characters load in the background, so laps switch without a loading screen.
  • Characters and props were built in Blender, Tripo and Mixamo, and animated with AnimationMixer.
  • A custom ShaderPass gives it the CRT / camcorder look.
  • There’s an online leaderboard. The server checks every lap, so posting a fake time with curl won’t work :slight_smile:

Initial load is about 17–20 MB, so it may take a few seconds on mobile data.

The music is an original track by SEVEN, the main character: “Click-Clack Snap feat. OCHIYO”.

I’d love feedback, especially on performance on your devices and whether the anomalies feel fair. Happy to answer questions about any part of it!

Loading the next encounter while you walk the current lap is the right call, and it leaves one hitch a loader can’t hide: the GLBs arrive early, but their materials still compile on the first frame they’re drawn. Warming programs with renderer.compileAsync before they’re needed took about 0.9 s off opening a new object in my scene, since KHR_parallel_shader_compile lets the driver compile without blocking. One trap with a ShaderPass in the chain: the scene renders into the composer’s render target, and output color space and tone mapping are part of the program, so compile with that target bound or you warm a different program. Do you notice a stutter right as a lap swaps on phones?

1 Like

Thank you so much for taking the time to look at this so carefully. I really appreciate it.

You were right about the trap. I was already calling compileAsync on each preloaded GLB, but with no render target bound, while the scene actually renders into the EffectComposer’s target. I checked the r170 source, and the warmed programs had AgX tone mapping and sRGB output, which the real ones don’t. I’ve fixed that so the composer’s target is bound first.

Your comment also made me measure things properly, and I found a few more causes:

  • I was warming the GLBs before swapping in my custom materials, so the warm-up compiled programs that never got used.
  • Some objects stay hidden until their encounter, so compile() skipped them.
  • Spot light counts change in the middle of a few encounters, which forces new variants.

On my desktop, with the browser’s shader cache defeated, the worst lap swap drops one or two frames (about 40 ms). I haven’t measured phones yet. That’s next, and I’ll post the numbers here.

Thanks again. This was exactly the kind of feedback I was hoping for!

Quick follow-up with the iPhone numbers I promised (iOS 18.7, shader cache defeated, about 5 minutes per run):

Before: a ~3 s freeze at the start, and 150–230 ms hitches on six lap swaps.
After: no shader hitches at the start or during play.

Your target tip was the starting point. The other fixes were warming the final swapped-in materials (including hidden ones), fixing the light count up front, and spreading the warm-up over frames, since one big compileAsync batch still blocked on the iPhone.

Thanks again, this made a real difference!

Going from a ~3 s freeze to no hitches on an iPhone is a great result, and thanks for posting the numbers. The hidden objects are the one I wouldn’t have guessed: compile() walking only visible meshes means anything waiting for its encounter is silently left out. Measuring with the shader cache defeated is the honest worst case too, since a returning visitor hides all of it. How did you split the warm-up across frames, a fixed number of materials per frame or a time budget per frame?

I think one thing worth considering is how players understand their mistakes after choosing the wrong direction.

Since the game requires 9 correct decisions in a row, I imagine it could get a little frustrating if someone keeps restarting without knowing which anomaly they missed.

Maybe an optional replay or a small hint after a failed run would help. Not during the actual game, of course, because that would ruin the tension. But once the run is over, being able to look back at the last encounter could make the learning process feel fairer.

I feel like this would be especially useful on mobile, where some visual details might be harder to notice on a smaller screen.

Would you consider something like that, or is figuring everything out through repeated attempts part of the experience you’re aiming for?

The hidden-material point is the useful catch here: a compile pass that only sees what is currently visible will always miss whatever comes next, so warming every variant up front is the only honest baseline. On the split, a time budget per frame tends to behave better than a fixed material count, because a heavy material can blow through a batch on weaker hardware regardless of how few you queue. To the other question about post-run feedback: showing the missed anomaly after a failed run, never during, would soften mobile restarts without touching the tension.

Thanks! And yes, the hidden objects surprised me too.

I used a time budget per frame, about 4 ms. Every mesh goes into a queue, and each frame I compile from it until the 4 ms is used up. Materials that are already warmed get skipped, and texture uploads go through the same queue. While the briefing screen is up, I hold off drawing the scene until the queue is empty, so the first visible frame doesn’t compile anything.`

A fixed count would have been hard to tune. On the iPhone, a single batch of about 20 programs took around 240 ms.

Thanks for the thoughtful suggestion! The silent reset is intentional. Like The Exit 8, the doubt (“was that an anomaly, or am I imagining it?”) is the core of the game, and telling players which one they missed would turn later runs into checking answers. Deaths already show the cause and a small hint, so it isn’t completely silent.

Your point about mobile is a fair one, though. If an anomaly is simply too small to see on a phone, that’s unfair rather than hard. The game already logs which encounter people fail on, so I’ll compare phone and desktop and fix any encounter that only phone players keep missing.

Thanks! That matches what I saw. I went with a time budget per frame, and a fixed count would have been hard to tune: on the iPhone, one batch of about 20 programs took around 240 ms.

On the feedback side, I’m keeping the reset silent during play, but I’ll check the failure data for anomalies that are hard to see on phones and fix those.