AI Exploration – Turning 2D Images into 3D Prints
01
Goal
Breaking the Flatness
Generative AI is making it easier than ever to create quick 3D objects from simple 2D images. Whether you’re exploring XR, 3D rendering, or printing, you can now turn images into printable 3D models. To prove the point, I transformed a pain-stricken stock photo star, into a desk-worthy gadget.
02
From 2D to 3D
New AI Models on the Block
In early 2025, two powerful models were launched that convert images into 3D objects without requiring manual sculpting:
Hunyuan3D by Tencent – Converts both images and text prompts into 3D objects (Restricted license for EU users)
TRELLIS by Microsoft – Works with images only but improves precision when given multiple inputs. (Permissive MIT License)
Both offer code samples for local inference if you have strong hardware, or you can run them in the cloud on platforms like Hugging Face or Runpod. Each model can generate both plain 3D meshes and textured (colored) versions. While both provide different hyperparameter controls, neither offered drastic improvements in my tests.
03
Trial and Error
Choosing the Right Model
Since different inputs yield different results, I tested both models to see which would give Harold the most respectable PETG reincarnation.
- Hunyuan3D had decent outputs, but TRELLIS was slightly better for my case—though it did introduce some 3D inconsistencies.
- I needed a high-poly result (4096 polygons), which wouldn’t run on Hugging Face’s ZEROgpu setup. Instead, I launched Runpod’s pre-configured container on an RTX A6000 with 48GB VRAM.
04
Editing
Making Harold 3D-Printer-Friendly
3D printing is an unforgiving medium, so I had to check the model’s integrity. Tools like Trimesh flagged the output as not strictly watertight, but my slicing software was much more forgiving. With a little cleanup in OpenSCAD, I made some tweaks, upgraded Harold’s mug, and re-exported the STL. The final print? Passable. Not flawless, but good enough for an experimental run. The rough print quality was more a result of my consumer-grade printer and the fast print mode I used than the AI-generated model itself.
05
Editing
The Bottom Line
These new 2D-to-3D models are exciting and functional, but they’re not magic. You’ll still need some hands-on tweaking, especially if you want clean, print-ready results. Want to try it out? Both models are free to test on Hugging Face.
Latest blog posts
Hospital VR Prototype
Make future innovation tangible with Virtual Reality