Generative AI Image-to-3D Services & APIs — a visual Benchmark

Michael T. Wagner
3 min readDec 13, 2024

--

Generative AI technologies have made remarkable strides in converting 2D images into 3D models, with innovative tools and features emerging at an impressive pace. To better understand the current capabilities of these services and APIs, I tested them all using the same image to see how their outputs compare.

Currently my favorite pick is Hunyuan3D-2 and Trellis, followed by meshy — but see for yourself.

The test image is a picture of a milling machine — the ABENE VHF-680, which I found in this PDF product brochure: VHF680CNC ENG 2011.

ABENE VHF-680

This machine presents a challenging test case due to its intricate geometry, including transparent components and protruding elements like the HMI. As a reference, I downloaded a traditional 3D model from Sketchfab to serve as ground truth:

ABENE VHF-680 Model

Cutter Machine — Download Free 3D model by Francesco Coldesina (@topfrank2013) [ea44a8b]

meshy.ai

(https://www.meshy.ai/?utm_source=E7PZL9)

Meshy is an easy to use web-service which offers a broad range of generative AI solutions.

It’s a paid service, everything costs coins, e.g. create the model, create the texture etc. expect to pay around 1$ per model until it’s finished.

Results are quite good, you can also add animation by text.

Result:

Shutterstock Gen3D API + Blender

(https://www.shutterstock.com/de/discover/generative-ai-3d)

Shutterstock gen3d API is currently only a beta - there is a standalone client and plugins for Blender and Maya. I’ve tested with the Blender Plugin :

https://generative-static.shutterstock.com/apps/shutterstock_gen3d_blender_1_1_3.zip

Documentation | Shutterstock

The result is not really usable — given the 25$ minimum subscription it’s rather disappointing.

Result:

Microsoft TRELLIS 3D

Microsoft TRELLIS is so far the best solution if you own a RTX A6000 graphics card because you can run it locally without additional costs and the resulting models are superb.

Result:

TRELLIS model

Stability.ai : stable-fast-3d

I just came across the stable-fast-3d solution — it’s fast but the results are better than shutterstock but not as good as meshy and TRELLIS.

Result:

stable-fast-3d model

DEEMOS HYPER3D- Rodin (Gen-1.5 V1.0)

https://hyperhuman.top/r/T069HNF4

Controllable Large-scale Generative Model
for Creating High-quality 3D Assets

Solid results so far — there are lots of options which I haven’t tried yet.

Result of Rodin

Tencent Hunyuan3D-2

Hunyuan3D-2.0 — a Hugging Face Space by tencent

--

--

Michael T. Wagner
Michael T. Wagner

Written by Michael T. Wagner

CTO and Co-Founder @ipolog.ai & synctwin.ai, creating clever solutions for smart factory

No responses yet