We talk to CTO Qixuan (QX) Zhang.
At SIGGRAPH Asia 2025 in Hong Kong, befores & afters got to sit down with Hyper3D.ai CTO Qixuan (QX) Zhang. QX breaks down everything you need to know about Hyper3D.ai and where it came from, which includes an interesting history in 2D image generation, 3D scanning and 3D asset generation.
We started with a discussion about the origins of Deemos, which is the overall company that includes Hyper3D.ai and tools like ChatAvatar, Rodin and OmniCraft.
b&a: Tell me about the origins of the company?
QX Zhang: We started this company in 2020. Our first product was a 2D image generator. At that time we didn’t have a diffusion model. The only thing we had was GAN, and we were using GAN to generate images, based on human hand-drawn sketches. We published an app called WAND on the Apple Store. The app became an overnight sensation, instantly climbing to #1 in the Graphics & Design category on the App Store in both China and Japan. Within just two weeks, we surpassed 1.6 million users.
After we did this project, we thought GAN could do image generation very well. So we turned to the 3D industry. The second project we worked on was 3D scanning. We built a Light Stage-like system to capture the human faces, including the geometry and the materials at sub-micron scales, and provided services to movies and games. We captured about 2000 different actors and actresses. Based on that data we decided to work on 3D character generation. This was at the end of 2023.
Then we attended our first SIGGRAPH, and we showed the demo for ChatAvatar at Real-time Live! We launched that after ChatGPT came out, but we started it before ChatGPT. In it, you can use text and images to generate 3D characters with PBR textures. You can then animate and render the character and put it into a game or movie.
After that project, we noticed that our 3D scanning service was decreased because a lot of people were directly using that product to generate characters, rather than scanning. So we decided to take it a step further into 3D generation, not only for characters, but also a whole range of other comprehensive digital assets. We spent about one year training our own 3D large language model. It’s a new architecture. We almost used the last of our money to train the model…and luckily we succeeded!
We published a paper to SIGGRAPH 2024 and we won the SIGGRAPH best paper honorable mention that year [for CLAY — a Controllable Large-scale Generative Model for Creating High-quality 3D Assets, awarded Best Paper Honorable Mention at SIGGRAPH 2024: https://dl.acm.org/doi/10.1145/3658146]. At the same year, we launched Rodin. It first showed up at GDC and then SIGGRAPH.
b&a: What was the main focus or breakthrough related to that research?
QX Zhang: That paper was CLAY. And before CLAY, people were working on 3D generations but basically using 2D image generators to generate 3D things. It would be using score distillation or generate multiple views and reconstruction to 3D again. But, we have been working on 3D reconstruction for a long time. We built the capture system, and we are familiar with 3D geometry. We know that if you use 2D images to recover the 3D surface, it will not be that good. You always need to do the cleaning up and re-topology.
So, when we started the project about 3D generation, we chose the 3D native pipeline. But before us, people were thinking that 3D didn’t have that amount of data for training because 2D image generation uses more than 7 billion images to train the model. But 3D only had 500,000. It was a small amount. People did not think it would be a success. So we designed our own. We followed a 3D implementation named 3DShape2VecSet, and based on that representation, we trained our own latent diffusion model. The quality is super good. The surface is smooth and clean.
#SIGGRAPH2024
🏆Best Paper Honorable Mentionedhttps://t.co/nKH8OIbHN2Introducing the magic behind #Rodin (https://t.co/sl0OCm9shs), CLAY: A Controllable Large-scale Generative Model for Creating High-quality 3D Assets
(1/n) CLAY is 3D #stablediffusion + #ControlNet + #LoRA. pic.twitter.com/w6Th9Qr5eZ
— Hyper3D by Deemos (@DeemosTech) June 24, 2024
For more details, please refer to our paper: https://arxiv.org/abs/2406.13897
b&a: What is the interaction now between Deemos, Hyper3D and Rodin?
QX Zhang: Deemos is the name of the company. We have a platform named hyper3D.ai. It’s also our website. Under the website, there are a lot of different kinds of products. For example, ChatAvatar, the character generation, and Rodin, the 3D asset generation. Rodin is the flagship model we are introducing to the community. There’s also OmniCraft, which is a 3D toolset.
b&a: What are users now using Hyper3D and Rodin for? What are you seeing out there?
QX Zhang: A lot of users are game developers. They are using Rodin to prototype for their games. It can save a lot of time. They used to find things in the 3D asset library, but now they can directly use Rodin to generate the asset and put it into their game. It’s not that game-ready, but you can use it as a prototype.
Designers are another major user group. They’re using Rodin to design different products. For example, car designs, and computer designs. They’re using Rodin for prototyping. For 3D printing, users can directly use Rodin to generate a 3D model and print it.
b&a: Is there a goal to make the models more game-ready?
QX Zhang: Yes, of course. Gaming is our primary focus. Right now, we are also working on AI topology and AI UV unwrap and textures, rather than just high detailed geometry generation.
b&a: How does a user right now jump onto hyper3D.ai and get started?
QX Zhang: You just drag and drop an image and click generate. One important feature we have is that hyper3D.ai is totally free to generate. You can generate unlimited assets for free. You only need to pay when you think the results are good and you want to download it. So, it’s free to use, it’s free to generate, but it’s paid to download.
Try it yourself: Visit hyper3D.ai and use code bnaRodinEdit for 14 days of complimentary Creator Plan access.
b&a: This is all changing so quickly. What is it like to keep up with changes in AI, and what are some of the things that you think will be coming in the near future with your tools, and generally?
QX Zhang: In 2026, our main focus is on the 3D editing, which means you can upload your own model to our website and edit it with AI tools. Rodin is mainly designed for designers as a tool, which means it’s not magic. We want the artists to be able to control every point, including using Rodin to edit their previous work, rather than just using Rodin as a magic tool to generate an asset from an image. We want them to be editable and let users modify assets in Rodin.
I think that’s one trend in 3D generation. And then there’s vibe coding. A lot of users are talking about vibe coding. They’re using large language models to generate code to make prototypes for their products. They don’t need the skills for web developing, instead they can use vibe coding to get a quick prototype of the product. I think that’s what will be happening next year in the 3D industry.
Brought to you by Hyper3D.ai:
This article is part of the befores & afters VFX Insight series. If you’d like to promote your VFX/animation/CG tech or service, you can find out more about the VFX Insight series here.
