Forge

Today I was trying to get my renders to stop looking like flat plastic.

I assumed my newest changes had broken something, so I started eliminating suspects. Reverted the latest feature. No change. Turned off the optimization. No change. Rolled everything back 10+ commits, to before any of it existed.

Still broken.

If reverting 10 commits changes nothing, the bug probably isn't in the commits.

Turns out the system responsible for making surfaces look realistic had never actually been switched on in my test setup. Everything silently fell back to a simpler rendering path, so I was testing code that couldn't affect the output.

The fix was one line.

After that, I got my first properly realistic render — and then got a door in the scene to open on command, rotating exactly 90°.

What I'm taking from it:

1. If a change does nothing, first check that the code path is actually running.

2. Write the startup checklist before you need it.

3. Silent fallbacks are some of the most expensive bugs to debug.

I also found another fun one: a generated scene came out completely black because it had zero light sources. It looked exactly like a camera/rendering bug.

Next up: fixing that and pushing the visuals further toward real.

What's the dumbest-in-hindsight bug that ate your week?

I've been building Forge, an AI engine that turns natural-language descriptions into explorable, interactive 3D environments.

The basic idea is simple:

Instead of describing something and getting back an image or a single 3D asset, you describe an environment and Forge tries to create a space you can actually enter, explore, and interact with.

For example, the goal is to go from something like:

"A small stone house surrounded by a garden, with a pond behind it."

to an actual explorable environment rather than just a picture of one.

I'm building this as a solo developer, and I'm still very early.

Right now I'm focused on small environments and the core technology needed to make them coherent, interactive, and visually convincing. There is still a lot to solve, especially around visual quality, performance, generation reliability, and the range of environments Forge can produce.

The interesting part for me isn't simply generating 3D objects.

I'm trying to build a system where natural language can describe a world, the system can reason about what that world should contain, and the resulting environment remains something you can actually explore and interact with.

The long-term vision is much bigger than the current MVP, but I'm deliberately starting small and testing whether the core idea is useful.

I'm using Indie Hackers to document the process honestly — the technical problems, experiments, failures, performance improvements, and eventually what real users think of it.

I'm especially interested in hearing from indie game developers, 3D creators, technical artists, procedural-generation enthusiasts, and anyone who has dealt with the pain of manually building environments.

If you could describe a 3D environment instead of building it piece by piece, what would you want to create?