Ranking AI 3D Generators: The Crafiq 3D Arena

Published on April 2, 2026

3DRankings

The AI 3D generation space has moved fast. A year ago, getting a clean textured mesh from a single image was a real challenge. Now there are many more models doing it with reasonable results, each claiming to be the best, each with cherry-picked demo images to back it up.

The only honest way to compare them is to run all of them on the same inputs and let people judge side by side without knowing which model produced which result. That is what the Crafiq 3D Arena is for.

The format is inspired by arena-style leaderboards like those at Artificial Analysis for image and video models, adapted for image-to-3D generation. You pick between two anonymous outputs, your vote is recorded, and that feeds into the global leaderboard at /rankings/3d.

How voting works

Each round shows you the input image and the input text prompt. On the left and right you get an interactive 3D viewer. You can rotate the model, zoom in, and toggle the texture off to look at the underlying geometry on its own. Neither model is labeled until after you vote.

That anonymity is the whole point. People vote based on what they actually see, not on brand recognition or which version number sounds more impressive.

What is being tested

The arena currently covers Tencent Hunyuan3D 3.1, Meshy 6, Microsoft TRELLIS.2, Tripo 2.5, ByteDance Seed3D, along with others. All models are tested against the same set of input images. Some of them can optionally take a text prompt alongside the image, and that is used where supported. The full set of test images and all generated 3D outputs are available at huggingface.co/datasets/Crafiq/3d-arena.

Below is one of the test cases in action. The task was to generate a 3D model of the bust of Agent Smith from The Matrix, including his signature earpiece and cable. The clip shows how Hunyuan3D 3.1, Meshy 6 and TRELLIS.2 each handled it.

None of them got the earpiece right. That specific geometry, a thin cable looping from a near-invisible earpiece, is not the kind of thing any of these models would have seen much of in training data. The overall shape and surface texture differ noticeably between models, but it is also the small specifics like this that tend to separate the better results from the weaker ones.

Where things stand

The arena is still early in terms of vote count, so the scores will shift as more people vote. That said, the current data is consistent enough to show a front-runner: Hunyuan3D 3.1 is out ahead for now.

Rankings are computed using an Elo-based system. All votes are replayed from scratch each time scores are recalculated, so early votes never permanently skew the results and new models can find their true position as more comparisons come in.

Try it yourself

You can vote on the arena yourself and see both the global leaderboard and your own personal rankings, which tracks your individual voting history and reveals your own preferences across models. If you want to actually generate 3D models, Hunyuan3D 3.1 and most of the other models in the arena are available to use directly in Crafiq.

If you want to check the current standings, see the leaderboard at /rankings/3d.