Custom 3D asset of my professor…kudos if you know the reference 🏄‍♂️

Custom 3D asset of my professor…kudos if you know the reference 🏄‍♂️

<aside> 🔥

*Full-stack React app built with Three.js. Uses TRELLIS (image-to-3D), text-to-image generation, and a promptable LLM agent to build dungeons from scratch. Built in a few weeks as part of a computer vision project.

Oh… and you bet all the assets shown throughout these demos were built in app.*

</aside>

App Tour 👟


dungeoncoder_overview.gif

Custom 3D Models 🧑‍🍳


This part is largely built around TRELLIS (image-to-3D) and features several different methods of generating the input image.

Designed so the user has a lot of creative control to generate what they want quickly!

🤖 For these demos, TRELLIS was hosted on a rented GPU (L40 from vast.ai) with a custom Flask inference endpoint to handle requests and share partial updates.

You can go from idea → image → asset in < 1 minute ⚡

You can go from idea → image → asset in < 1 minute ⚡

Image editing in action. Can edit uploaded images, existing 3D models, or combine multiple existing models 😎

Image editing in action. Can edit uploaded images, existing 3D models, or combine multiple existing models 😎

Couple of the default assets the dungeon design agent has access to. All generated through the app ofc 💥

Couple of the default assets the dungeon design agent has access to. All generated through the app ofc 💥

How to Train Build Your Dragon Dungeon


I’ll focus mostly on the agential system. Feel free to dive into some of the weeds if you so insist…

The Weeds: Why Scene Generation is Hard

LLM Interior Designers: Just Becoming Good Enough

TLDR; to do real-time scene generation that is consistent, high-quality, and interesting — you can either use LLMs (only recently became good enough in early 2025!) or you can spend tons of engineering man-hours building a custom system with manual rulesets (not the scope…).