Table of Contents
Overview
We have already looked at GenAI coding, and pixel art generation. I thought it might be a good time to look at 3D Game assets, to see what generative ai can make. Is it any good? How does it compare to what humans create? And as a record of how far and fast AI improves in 3D asset generation in future
I will be evaluating three 3D Asset generation platforms
- Tripo – commercial GenAI, English, Tripo frontend
- Meshy – commercial GenAI, English, Meshy frontend
- Hunyuan3D-2.1 – Open Weight Model, Chinese, HuggingFace frontend
- Hunyuan3D-2.5 (7th July 2025) – Open Weight Model, Chinese, 3d.hunyuan.tencent.com frontend
The Output
I’ll jump straight in here with the output from the GenAI platforms. Afterwards there is an Appendix with more background detail.
Tripo – via studio.tripo3d.a
URL: https://studio.tripo3d.ai
Rating: 5/10 ★★★★★☆☆☆☆☆
Vertices: 272,857 (!) – simplification step got errors
Triangles: 499,837 (!) – simplification step got errors
Filesize: 11.57 MB – Download here
Meshy – via meshy.ai/workspace
URL: https://www.meshy.ai/workspace
Rating: 7/10 ★★★★★★★☆☆☆
Vertices: 4,897
Triangles: 10,000 (listed as Faces, so please adjust appropriately)
Filesize: 2.9 MB – Download here
Trellis – via Hugging Face
URL: https://huggingface.co/spaces/trellis-community/TRELLIS
Rating: 7/10 ★★★★★★★☆☆☆
Vertices: 21,703
Triangles: 26,423
Filesize: 2.0 MB – Download here
Hunyaun3d-2.1 – via Huggingface.co
URL: https://huggingface.co/spaces/tencent/Hunyuan3D-2.1
Rating: 6/10 ★★★★★★☆☆☆☆
Vertices: 28,267
Triangles: 40,000
Filesize: 3.59 MB – Download here
Hunyaun3d-2.5 – 1/2) lowpoly
URL: https://3d.hunyuan.tencent.com/lowpoly
Rating: 7/10 ★★★★★★★☆☆☆
Vertices: 16,912
Triangles: 15,446
Filesize: 9.49 MB – Download here
Hunyaun3d-2.5 – 2/2) highpoly
URL: https://3d.hunyuan.tencent.com/
Rating: 6/10 ★★★★★★☆☆☆☆
Vertices: 334,590 (!)
Triangles: 499,944 (!)
Filesize: 26.2 MB (!) – Download here
Conclusions and Practical Implications
The Good: 3D assets magicked up by GenAI.
The Bad: The tools themselves have some bugs in places. The output is suitable if viewed from far away. Anywhere close up and it looks bad. The text on the apple cart sign in all cases ended up as gibberish. This clearly needs to be improved.
The Ugly: Optimisation is poor, textures ‘leak’… like on the apples and fruit. UV Mappings look messy (although I must admit I am not an expert in just how important this ‘messy UV Mappings’ is).
🏆Best Generative AI – 3D Video Game Assets (July 2025):🏆
🏆Trellis – 7/10 – ★★★★★★★☆☆☆
🏆Meshy.AI (commercial) – 7/10 – ★★★★★★★☆☆☆
🏆Hunyuan-3D 2.5 Low Poly – 7/10 – ★★★★★★★☆☆☆
All of these created actually low poly models. None is perfect and I expressed my doubts how practical this output is for using in professional video games. Despite this we are clearly heading towards even better results in future, and perhaps there are many usecases where these results would be acceptable.
Trellis had the best looking fruit on the cart, so it is listed first.
Appendix – Technical and Sciency Stuff
Subject Matter
I decided to use a subject I’ve previously used in 2D comparisons “A fruitcart. The sign on the cart reads ‘Apple-a-day'”
This example tests 3D modeling and also text labels.
Workflow Step #1 – Text to Image
I started with the commercial offering Tripo.
Tripo took the text words and created a preview image. After that preview is made it generates a 3D model from the 2D preview image. You are free to supply your own image. I was keen to see what the text portion did so that it was end-to-end GenAI.
I generated a couple candidates for "A fruit cart. The fruit cart sign reads 'Apple-a-day'". Below are the images and I marked which one I chose to proceed with and which I ignored. I chose the model that had the nicest sign portion and the better variation of fruit colours.
As you can see it looks as though Tripo is using a 2D text to image genai to make the candidate previews. If I was to guess (I have no inside knowledge) it looks a lot like ChatGPT ImageGen 4.1.
I used the image made by Tripo for all the 3D asset creation even on non-Tripo platforms. This keeps some small amount of consistency.
Workflow Step #2 – 2D image to 3D Asset
About the Ideal 3D Model
In this article we are looking to make low-poly (low polygon count) models. During my testing this was not always possible due to internal errors with some of the platforms. So unfortunately for the error platforms we were left with high-poly (very high polygon count) 3D models. Games want low polygon count models which are easier for the customers’ GPU to draw quickly. High Poly assets are not used in games, although sometimes in non-game settings like 3D CAD. Also there are external High Poly -> Low Poly conversion tools so I figure that High Poly is okay for the gist.
UV Mapping
UV Mapping is how the 3D model (3D) knows what portions of the 2D Texture (aka Material (also 2D)) to display on what parts. In professional 3D asset creation this is an important part of the overall quality of the asset. Well UV Mapped assets are easier for teams of artists to work with when creating the 2D texture. Important factors are that portions which are large contiguous regions on the model are represented by similar large contiguous regions on the texture. This helps the artists understand where to make changes without the model.
What POH Human Professionals make
Below is a 3D asset made by a Plain Old Human (POH) Professional. The Model and UV Mappings are clean and well structured. It is easy to see which parts of the 3D model map onto the texture
Here is the UV Mapping – You can clearly identify by eye which 2D regions correlate with the 3D model. The regions are contiguous and do not have missing parts. The shapes are unwrapped like cardboard boxes unfurled into its original shape. Above I highlighted Wheels and the Chassis. Can you spot which area on the UV mappings are likely to have the Wheel and Chassis on the texture (below)
And here is the POH human professional texture material. It’s not pretty, but you can identify the sections
GenAI UV Mappings
I have chosen the best of the bunch tested above, the Meshy.AI, Hunyuan3D-2.5 low poly result. The others are much worse and not shown.
The model looks low poly to the eye
When you look at the UV Mapping you get a shock… There are hundreds(!) of little UV sections. These all have to be painted separately. Any hand edits are all over the (UV) Map. Maybe some professional 3D modellers can comment about whether they would be happy getting an asset that looked similar to this?
And below is the texture from GenAI… What a confusing mess. The fruit is all over the place. Ideally all the fruit would be UV mapped into roughly the same section of the 2D material
