Vi3W
Skip to content

Guides

The top 3D AI tools leaderboard: what 3D Arena's Elo ranking shows, and what it misses

The 3D Arena leaderboard for image-to-3D AI: the November 2026 top 10, how its Elo ratings work, and what a vote-based ranking can't tell you.

5 min read

A three-step glowing podium in a dark arena holding a lantern, a robot and a dragon, with spotlights crossing above
Illustration. A leaderboard shows which results people preferred, not which tool suits your job.

Quick answer

Which 3D AI tools top the leaderboard?

On 3D Arena, a public vote-based ranking of image-to-3D models, PicGen3D led on November 21, 2026, followed by TRELLIS-3DGS and Meshy-6. It ranks by Elo from head-to-head votes. Tripo, Rodin and Hunyuan 3.x weren't listed, and newer entries have few votes, so treat it as one signal.

At a glance

3D Arena leaderboard snapshot (24 models, 114,236 votes), from the 3D Arena Space and each project's own pages, checked November 21, 2026.
RankModelEloOpen weights
1PicGen3D1415No
2TRELLIS-3DGS1398Yes
3Meshy-61348No
4404_GEN1334unknown
5Meshy-71326 (only 324 votes)No
6TRELLIS1303Yes, MIT
7Zaohaowu3D1295unknown
8Hunyuan3D-21283Yes, Tencent community licence
9Hunyuan3D-2.11271Yes, Tencent community licence
10InstantMesh1269Yes, Apache-2.0
11SAM-3D-Objects-3DGS1268Gated, SAM License
18TRELLIS.2-4B1129Yes, MIT

What 3D Arena is

3D Arena is a public leaderboard for image-to-3D models, hosted as a Hugging Face Space (3d-arena/3d-arena, MIT-licensed, built by Dylan Ebert). It works like a taste test. You're shown two 3D results generated from the same input image, without being told which model made which, and you vote for the one you prefer. Thousands of those votes add up to a ranking.

The method is described in a paper (arXiv 2506.18787). In mid-2025 it reported 123,243 votes from 8,096 users across 19 models. On November 21, 2026, the live leaderboard listed 24 models and 114,236 votes. The two counts were taken at different times and may be counted differently, so don't read much into the gap.

It is the most-cited public ranking of 3D generators, which is exactly why it's worth understanding before you lean on it. For our own model-by-model comparison, see the best AI 3D generation models for 2027.

How Elo ratings work

3D Arena uses Elo, the rating system from chess. Every model starts at 1200. After each vote, the winner takes points from the loser. How many depends on what was expected:

  • If a highly rated model beats a low-rated one, that was expected, so few points move.
  • If a low-rated model beats a high-rated one, that's a surprise, so many points move.

Over thousands of votes, ratings settle where each model wins about as often as its rating predicts. The gap between two ratings is what matters, not the absolute number. In standard Elo, a 100-point gap means the higher-rated model is expected to win about 64% of head-to-head votes, and a 200-point gap about 76%.

Read the table with that in mind. PicGen3D (1415) and TRELLIS-3DGS (1398) are 17 points apart, which is close to a coin flip. PicGen3D and InstantMesh (1269) are 146 points apart, which is a real preference.

A long dark shelf of small 3D objects including a teapot, a robot toy, a chair, a dragon and a lantern
Illustration. Arena votes compare results object by object, then average them out.

What the November 2026 ranking shows

  • An unfamiliar name at the top. PicGen3D, a closed model, led with 1415.
  • Open models compete. TRELLIS-3DGS was second, the original TRELLIS sixth, and Hunyuan3D-2 and 2.1 eighth and ninth. InstantMesh, from 2024, was still tenth.
  • Meshy is the top commercial vendor listed. Meshy-6 was third. Meshy-7 was fifth but had only 324 votes.
  • Newer isn't always preferred. TRELLIS.2-4B, Microsoft's 2025 successor, sat 18th at 1129, well below the original TRELLIS. Meshy-5 was 23rd.

Why an Elo rating depends on the field

Elo has quirks that matter when you compare 3D models.

  • A rating is relative to the other entries. Add a batch of strong new models and everyone else's rating can drift down without anything about them changing. A 1300 today doesn't mean the same as a 1300 a year ago.
  • Opponents aren't equal. A model that mostly met weak opponents early can carry a flattering rating until it meets the leaders.
  • Inputs vary. Some objects suit some models. If a model happened to be shown more of the objects it handles well, its rating benefits.
  • Ties of taste. Voters don't share one standard. Some prefer sharp texture, some clean silhouettes. Elo averages those tastes; it doesn't tell you which one matches yours.

None of this makes the ranking useless. It makes it a measure of general preference across many objects and many voters, which is a different thing from "best for my project".

What the leaderboard can't tell you

It isn't the whole market. Tripo's current models, Hyper3D Rodin and Tencent's HY 3D 3.x weren't on the board. A tool that isn't listed hasn't lost; it hasn't been entered.

Low vote counts move fast. A new entry with a few hundred votes, like Meshy-7, can rise or fall a lot as votes arrive. Trust ratings built on thousands of votes more.

Votes judge looks, not usability. A vote records which result someone preferred in the viewer. It doesn't record whether the mesh is quads or triangle soup, whether the UVs are usable, what the face count is, or whether it will rig. For a game or a print, those can matter more than looks.

Splats and meshes are mixed. TRELLIS-3DGS and SAM-3D-Objects-3DGS output Gaussian splats, which can look very good in a viewer but aren't meshes you can edit, print or rig. If you need a mesh, compare mesh entries with each other. See Gaussian splatting vs meshes.

It's image-to-3D only. Text-to-3D, rigging, retexturing and retopology aren't ranked.

It's a snapshot. Models churn quickly. Check the date on any ranking you're quoting, including this one.

Other 3D leaderboards

  • 3D Worlds Arena (3d-arena/3d-world-arena), a newer sibling Space for world generators, updated November 13, 2026.
  • 3DGen-Arena (3DTopia), an older arena, last updated March 2025, so treat it as stale.
  • SA-3DAO leaderboard (Meta), a benchmark built around SAM 3D's artist-object dataset, from June 2026.

How to use a leaderboard well

Use it to build a shortlist, then test the shortlist on your own work. Pick three objects you'd actually make, run each through two or three models with a good rating and enough votes, and compare the files rather than the previews: wireframe, back side, face count, texture and file size. A model that ranks fifth overall can be first on the kind of object you care about.

Run your own arena

Vi3W runs 25 image-to-3D models, including Meshy, Tripo, Rodin, Hunyuan 3D and TRELLIS, on the same image, with TRELLIS free by default and free downloads, so you can compare the actual files. For the bigger picture, see the best 3D AI going into 2027. Open Vi3W.

Frequently asked questions

top 3d ai tools leaderboard

The best-known one is 3D Arena on Hugging Face. On November 21, 2026 its top five were PicGen3D, TRELLIS-3DGS, Meshy-6, 404_GEN and Meshy-7. It covers image-to-3D only and leaves out several major commercial models.

3d ai tools leaderboard

3D Arena ranks image-to-3D models by Elo from user votes between two outputs of the same image. A sibling 3D Worlds Arena covers world generators, and Meta runs a separate SA-3DAO benchmark leaderboard for its SAM 3D artist-object set.

best ai 3d generator reddit

Forum threads give you individual opinions on individual objects. A leaderboard like 3D Arena aggregates many such judgements, which is more stable but still only measures what voters can see. Use both as pointers, then test on your own image.

Last reviewed