Generating 3D models with AI tools gets a bit easier with not one, but two new options.
For decades the creation of complex 3D models has been the realm of those expertly trained in the use of advanced 3D CAD or sculpting software. It’s been a major barrier to the adoption of 3D printing, because it causes operators to seek others to do the design work for them.
With the introduction of increasingly powerful AI tools this barrier could soon come down, at least for some types of designs.
We’ve seen multiple attempts at creating tools that can generate 3D models. Some use a text prompt as input, some use 2D images as input and some a combination of both. One we recently looked at was Meshy.ai, which produces pretty decent output from simple input.
Now we’ve learned that two new and powerful 3D AI services have come online.
Tencent’s Hunyuan 3D Creation Engine
One is from China: Tencent’s Hunyuan 3D creation engine, which you can test at the Tencent Hunyuan 3D Global site. The Hunyuan 3D creation engine was previously available in China, but now global access has been enabled for individuals and companies.
The Tencent Hunyuan 3D Global site is free to access, only requiring an email to login. The site offers up to 20 free generations per day, which should be plenty for many users. There are corporate plans for more intensive use.
The Hunyuan 3D creation engine accepts text, image or even sketches as input, and quickly develops a full 3D model. Well, my test took 190 seconds, but that’s far faster than I could sculpt a similar model manually.
From my input image, I was able to generate this decent 3D model, which was fully textured:
However, if you turn off the textures and look at the geometry, it seems that the moustache and beard are simply “painted on”, rather than having actual geometry of their own.
Nevertheless, it’s a pretty good result from a free service, perhaps one of the best I’ve tried so far.
Meta’s SAM 3D
The other service is from Meta, their new SAM 3D technology. It’s comprised of two specifically trained AI models, SAM 3D Objects and SAM 3D Body. One is for arbitrary objects and the other is specifically tuned for the human body — faces and poses.
The models accept as input 2D images, which it can then transform into full 3D models. Of note is that these models also include Meta’s SAM 3 model, which understands images and videos. This likely makes the 3D model generation more accurate.
Meta provides a “playground” where you can fiddle around with these new models at no charge. I gave the SAM 3D Body model a test using the same image, but was a bit disappointed.
It turns out that SAM 3D Body is literally that: it generates full bodies, with two arms and two legs. By sending in a headshot, it gave me a default body pose with the head in the same orientation as the input image.
I tried the other model, SAM 3D Object, with the headshot to see how it worked. You must select all the regions in the headshot image to proceed. This is because the use case Meta is targeting is a general picture where you only want to model the toaster on that table — select it from the image. Here I’m trying to model an entire headshot.
But it mostly worked. Here you can see the output generated. It’s nowhere near as good as Huyuan’s, and is much more cartoonish in style.
I then tried SAM 3D Bodies with actual human bodies in an action sports shot. You can see that the AI was easily able to identify all the people in the image and correctly capture their poses. You can even spin around the image to see what they would look like from different angles.
Let’s try it on a more proper use case. I uploaded a scene with a chair, and attempted to extract a 3D model of one of the chairs. It seems to have worked extremely well:
There’s one catch, and it’s a big one: there doesn’t seem to be a way to download the resulting 3D model. You can export a video of it, but that’s not particularly helpful when you want to print your object. However, it’s pretty clear that this could be offered, it’s just that Meta seems to have decided not to in this demo.
One more thing: Meta has made these models open source, so that means they will no doubt begin to appear in new services providing the horsepower required to run them. Expect to see a lot more 3D model generation services soon.
The two new AI systems have slightly different purposes, but 3D printer operators now more options to create content.
