ORLA

ORLA turns a text prompt into a fully interactive 3D model. Type "a futuristic sports car," and ORLA generates it — first as an image, then as a real 3D asset you can rotate, zoom, and inspect in the browser.

We built the entire pipeline: the interface, the AI orchestration behind it, and the infrastructure to run two different AI models back to back without the user ever seeing the seams.

Going from "a sentence" to "a usable 3D model" isn't one AI call — it's a chain. Text has to become a clean 2D image first, and that image has to be shaped exactly right — isolated object, plain background, even lighting — before a 3D model can be extracted from it at all. Every step in that chain is slow, and none of it can block the interface while it runs.

The product needed to hide that entire chain behind something that feels instant: type, wait a moment, get a model you can actually interact with.

• Prompt-to-image generation — User prompts are enhanced with parameters tuned for 3D conversion, then run through Stable Diffusion 3.5 to produce a clean, isolated render.
• Image-to-3D conversion — The generated image is passed to Hunyuan3D-2, which builds a full 3D model from it.
• Interactive 3D viewer — The finished model renders directly in the browser with Three.js and React Three Fiber — orbit to rotate, scroll to zoom, right-click to pan.
• Multiple angles per generation — Users see several renders of their result, not just one flat image.
• Async task pipeline — Celery workers and a Redis queue handle the generation chain in the background, so the interface stays responsive while the AI models do the heavy lifting.

• React, TypeScript
• Three.js, React Three Fiber
• Tailwind CSS
• FastAPI
• Celery, Redis
• Stable Diffusion 3.5 (Stability AI)
• Hunyuan3D-2 (Tencent)
• Docker

• A single prompt produces a real, interactive 3D asset — no 3D modeling software, no manual sculpting.
• Two heavyweight AI models chained into one smooth, async pipeline instead of a slow, blocking process.
• A generation workflow built to run reliably at scale, containerized end-to-end with Docker — running locally, not yet deployed publicly.