Best Open-Source AI 3D Models, Measured (2026)

Only one open-source AI 3D model passes both tests that matter to a working creator. TRELLIS.2 is the only one of 17 open 3D model families that you can sell from under a plain MIT license and that runs in all four creator apps we checked. The most downloaded open 3D model inside ComfyUI is a different one: Tencent's Pixal3D, built on TRELLIS.2, whose ComfyUI files drew 806,174 downloads in the past 30 days. That is more than the other four ComfyUI 3D repackages combined.

We pulled data from five public sources: the Hugging Face Hub API for 17 open 3D model families (the official repos, the Comfy-Org repackages, and every community model that declares one of them as its base in the Hub's model tree), the license text of every family, each maker's own hardware figures, the code history of ComfyUI, and the source code of four creator apps that run 3D models: ComfyUI, Wan2GP, LocalAI and AMD's Lemonade. Here is what the numbers say about which open 3D model to use, which ones you can sell from, and which ones you can open in an app instead of a code editor.

Key Findings

1. TRELLIS.2 is the only model that clears every bar

Three questions decide whether an open 3D model is useful to someone making assets for a game, a client or a print shop. Can you sell what it makes? Can you run it in an app you already have? Does it give you a finished, textured model? TRELLIS.2, Microsoft's 4-billion-parameter image-to-3D model, is the only family that answers yes to all three.

Its weights are MIT licensed. It turns one image into a mesh with PBR materials, the color, roughness and metal maps that game engines and Blender expect. And it is the only family in all four apps we checked: ComfyUI, Wan2GP, LocalAI and Lemonade. Microsoft's card asks for a graphics card with at least 24 GB of memory and reports about 3 seconds for a 512-resolution model and about 17 seconds at 1024 on an NVIDIA H100, a data-center card. The apps bring that down: Wan2GP ships compressed 8-bit versions, and LocalAI's documentation puts its high-resolution 1024 pipeline at roughly 10 GB of video memory. Those figures are the makers', not ours; we did not generate models for this report.

It also has 1,378 likes on Hugging Face and 101 Spaces where you can try it in a browser, and it went from no creator app to four in under three months: Lemonade added it on 15 July, LocalAI on 29 July, ComfyUI on 21 August and Wan2GP on 10 October.

2. Pixal3D is the most downloaded open 3D model in ComfyUI

Pixal3D, from Tencent's ARC Lab, was released at the end of April, and ComfyUI's code added it as a built-in node on 21 August. In the past 30 days its ComfyUI repackage drew 806,174 downloads. The other four Comfy-Org 3D repackages drew 567,052 between them: TRELLIS.2 368,823, Hunyuan3D 2.1 93,420, Hunyuan3D 2.0 68,679 and SAM 3D Body 36,130.

Pixal3D is less a rival to TRELLIS.2 than an upgrade on top of it. It reuses TRELLIS.2's decoders, and the Pixal3D ComfyUI download includes TRELLIS.2's shape and texture files. What it changes is how the picture feeds the model: it projects each pixel of your image into 3D, so the shape and texture stay closer to the photo you started from. ComfyUI ships the two side by side in one template, "Pixal3D & TRELLIS.2: Image to Model". Like TRELLIS.2, it is MIT licensed. Its card offers a low-memory mode, which drops the default resolution from 1536 to 1024.

The community is also repackaging it for smaller cards without registering it in the model tree. The largest such upload, a compressed 8-bit Pixal3D, drew 59,703 downloads in the same 30 days.

3. Download counts are the wrong way to pick a 3D model

Look at Hugging Face's numbers alone and you would pick the original TRELLIS from December 2024. Its image model drew 2,308,737 downloads in the past 30 days and 41,646,047 since release, 34.6% of all the downloads we counted across 17 families. None of the four apps runs it. The Hub counts a download each time a script fetches certain model files, so demo Spaces, notebooks and test servers that load a model on every restart count every time.

The counter also misses things. For the official repos of Pixal3D, TripoSplat, Step1X-3D and UltraShape it records zero downloads, because their files are not the kind it counts. Pixal3D's official page shows zero while its ComfyUI copy shows 806,174. The repackages are built the same way as each other, so they can be compared with each other. For everything else, whether an app runs the model tells you more than the download number does.

4. MIT on the label, a non-commercial helper inside

Every open image-to-3D model first cuts your subject out of its background. TRELLIS.2's official pipeline file does that with RMBG-2.0, a background remover from Bria that is licensed CC BY-NC 4.0, non-commercial, with a paid license for business use. Pixal3D's official pipeline names the same model. TRELLIS.2's code calls it whenever your input image has no transparent background. TripoSG's official code uses Bria's older RMBG-1.4 for the same job. Wan2GP's developer noticed: its TripoSG code says it uses an open background remover "in place of BriaRMBG-1.4 (non-commercial)".

So "MIT" covers the 3D model, not every model the official script downloads. None of the four apps we checked references RMBG-2.0 in its code, so this applies to people running Microsoft's or Tencent's own scripts. The fix is simple: cut the object out yourself and give the model a PNG with a transparent background. The official TRELLIS.2 code then uses your cut-out as it is and never loads the background remover.

5. "Free to sell" means four different things

Every family here is called open, and the licenses split four ways. Nine let you sell what you make with nothing more than the license notice: TRELLIS.2, Pixal3D, TripoSG, TripoSplat, Direct3D-S2, TripoSR and the original TRELLIS (all MIT), and Step1X-3D and UltraShape (both Apache 2.0).

Three come with conditions. SAM 3D Objects and SAM 3D Body use Meta's SAM License: royalty-free and worldwide, but you request access to the weights first, and trade-control and military-use limits apply. Stable Fast 3D and SPAR3D use the Stability AI Community License: free for commercial use if you or your company make under $1 million a year, after you register at stability.ai/community-license.

Three depend on where you live. The Hunyuan3D 2.1 license, Hunyuan3D 2.0's and HY-World 2.0's all open with the same statement: the agreement "does not apply in the European Union, United Kingdom and South Korea." Section 5c goes further: you must not use the model's "Output or results" outside that territory. Tencent claims no rights in what you make, and companies with more than 1 million monthly users need a separate license. Pixal3D, also from Tencent, carries none of this: its own license is plain MIT.

Two rule out paid work. Lyra 2.0 ships under NVIDIA's Internal Scientific Research and Development Model License, and its card says you may not use it "for the purpose of generating works for sale or distribution." Apple's SHARP is licensed "exclusively for Research Purposes," which the license defines as non-commercial.

6. 3D has almost no remix culture yet

Across all 17 families, the Hub lists 110 community models built on them, and only 7 of those are LoRAs, the small add-on files that teach a model a style. Last week we measured open music models the same way, and one music model, ACE-Step 1.5, had 95 community models and 38 LoRAs on its own. Image models count their LoRAs in the thousands.

Most 3D community uploads are not new styles; they are the same model made to run somewhere else. Of TRELLIS.2's 10, three are compressed versions, including the GGUF build LocalAI installs. SHARP has 10, eight of them compressed or converted copies, several for Apple's Core ML. Only 23 community models appeared across all 17 families in the past 30 days, and TRELLIS.2 and SAM 3D Body led with 5 each. For a creator that means you cannot yet download a "low-poly" or "stylized character" add-on the way you can for an image model. The style has to come from the image you feed in.

7. Nine of 17 run in no app, and ComfyUI's arrival time ranges from 1 day to 297

We read the source code of four apps that creators use to run open models locally and checked which 3D models each one actually registers. ComfyUI runs six families: TRELLIS.2, Pixal3D, Hunyuan3D 2.0 and 2.1, TripoSplat and SAM 3D Body. Wan2GP, the one-window app built for modest graphics cards, runs five: TRELLIS.2, Hunyuan3D 2.1, TripoSG, TripoSplat and WorldMirror 2.0, the reconstruction part of HY-World 2.0. It added all of them in one update on 10 October. LocalAI and AMD's Lemonade each run one: TRELLIS.2. SwarmUI, which we also checked, runs no 3D model.

That leaves nine families in none of them: Step1X-3D, Direct3D-S2, UltraShape, SAM 3D Objects, Stable Fast 3D and SPAR3D, TripoSR, the original TRELLIS, Lyra 2.0 and SHARP. To use those you install them from code or try them in a browser demo.

Speed of arrival varies wildly. ComfyUI's code history shows TripoSplat landed one day after its weights went up. Hunyuan3D 2.0 took 58 days and 2.1 took 83. Pixal3D took 113. TRELLIS.2, the model that now runs everywhere, took 263 days, and SAM 3D Body 297. For most of the past year, the best open 3D model needed community plugins or its own install.

8. The open world models are the hardest to use for paid work

Three families here build whole scenes rather than single objects. HY-World 2.0 turns text, an image or a video into a 3D world of Gaussian splats and meshes. Lyra 2.0 builds an explorable splat world from one image. Apple's SHARP turns a single photo into a 3D splat scene "in less than a second on a standard GPU", according to Apple, and runs on CPU and Apple Silicon too.

None of them is free to sell from everywhere. Lyra 2.0 and SHARP are research-only, and HY-World 2.0 is excluded in the EU, UK and South Korea. Only one creator app runs any part of them: Wan2GP's WorldMirror 2.0, which reconstructs a scene from photos or video rather than generating one. HY-World 2.0 has 719 likes but one community model. If you want an open world generator you can use for client work, it does not exist yet in this group.

Trend Analysis

Rising: TRELLIS.2 and the models built on it. It went from zero creator apps to four between July and October: Lemonade and LocalAI ported it, ComfyUI made it a built-in node in August, and Wan2GP added it on 10 October. Pixal3D rides on the same decoders and leads ComfyUI downloads.

Stable: Hunyuan3D 2.1. It still has the largest community of the 2025 releases, 18 models, and runs in two apps, but Tencent has posted no newer version of the main Hunyuan3D generator on Hugging Face since 2.1 in June 2025, only spin-offs such as Hunyuan3D-Omni and Hunyuan3D-Part, and its license excludes three major markets.

Emerging: single-image Gaussian splats. TripoSplat (MIT) reached ComfyUI in a day and Wan2GP in October, and splats are what web viewers and AR apps can show straight away. Our guide to putting a photo-to-splat model on the web covers that path.

Declining: the 2024 generation. TripoSR, Stable Fast 3D and the original TRELLIS still draw large download numbers, but no app we checked runs them, and between them they gained one community model in the past 30 days.

Predictions

1. The next major open image-to-3D model will be built into ComfyUI within a week of release. TripoSplat took one day in June. The backlog of slow arrivals (TRELLIS.2 at 263 days) was cleared in August, and ComfyUI now has shared 3D nodes for loading, cleanup and export that a new model can plug into.

2. Pixal3D will be in at least two of the four apps by the end of January 2027. It shares TRELLIS.2's decoders, which all four apps already run, and it leads ComfyUI downloads by a wide margin.

3. 3D LoRAs will stay rare through mid-2027. Seven LoRAs across 17 families, and 23 community models of any kind in a month, say training add-ons for 3D models is still a specialist job. We will count again in six months.

4. A license with Tencent's EU, UK and South Korea exclusion will not be the one creators standardize on. Tencent's own lab chose MIT for Pixal3D, and Pixal3D already out-downloads both Hunyuan3D repackages in ComfyUI combined, 806,174 to 162,099.

What This Means for Creators

If you sell 3D assets, or make them for clients: start with TRELLIS.2 or Pixal3D in ComfyUI, both MIT. Remove the background yourself and feed in a transparent PNG; if you run Microsoft's or Tencent's own scripts, that keeps the non-commercial background remover out of your pipeline. Our photo-to-3D walkthrough in ComfyUI covers the steps.

If you live or sell in the EU, the UK or South Korea: skip Hunyuan3D 2.0, Hunyuan3D 2.1 and HY-World 2.0. Their licenses do not apply there, and they cover what you make with the models as well as the models themselves.

If you have a mid-range graphics card: Wan2GP is the easiest door. It runs five 3D families in one window and ships compressed versions. LocalAI and Lemonade run TRELLIS.2 behind a simple app and API, and Lemonade's TRELLIS.2 builds cover AMD cards as well as NVIDIA; we covered its 3D panel when it launched.

If you need posed characters: SAM 3D Body in ComfyUI turns a photo of a person into a posed body mesh, and ComfyUI can export it with a skeleton. Request access to Meta's weights first, or use the ComfyUI repackage, which carries the same SAM License.

If you make AR, web or virtual production scenes: TripoSplat gives you a splat from one image under MIT. Treat SHARP and Lyra 2.0 as previews of where scene generation is going, not as tools for paid work.

If you would rather not run anything locally: our comparison of Meshy, Tripo, Rodin and other hosted 3D generators covers the paid route.

Full Data

Downloads are the Hugging Face Hub's 30-day counts, which record requests for certain model files, so they favor models that scripts load often and record zero for some repos. Likes, community models and Spaces are summed across each family's official repos. Community models are repos that declare a family's official or Comfy-Org repo as their base. App support was checked in each app's source code on 21 October 2026, and ComfyUI arrival dates come from its code history. Memory and speed figures come from each maker's model card or code page and were not measured by us.

Frequently asked questions

What is the best open-source AI 3D model generator in 2026?

For most creators, TRELLIS.2. It is MIT licensed, turns one image into a textured mesh with PBR materials, and runs in ComfyUI, Wan2GP, LocalAI and Lemonade, the only family in all four. Pixal3D, built on TRELLIS.2 and also MIT, is the most downloaded open 3D model in ComfyUI and keeps closer to the input photo.

Can I sell 3D models made with TRELLIS.2?

Yes. TRELLIS.2's weights are MIT licensed. One catch: Microsoft's official pipeline removes the image background with Bria's RMBG-2.0, which is licensed for non-commercial use. Give TRELLIS.2 an image with a transparent background and the official code skips that step. The ComfyUI, Wan2GP, LocalAI and Lemonade versions do not reference RMBG-2.0.

Can I use Hunyuan3D 2.1 in the EU or UK?

No. The Tencent Hunyuan 3D 2.1 Community License states that it does not apply in the European Union, the United Kingdom or South Korea, and it forbids using the model or its output outside the licensed territory. The same exclusion appears in the Hunyuan3D 2.0 and HY-World 2.0 licenses. Pixal3D, also from Tencent, is MIT licensed with no territory limit.

Is Pixal3D free for commercial use?

Yes. Pixal3D's weights and code are released under the MIT License by Tencent. It reuses parts of TRELLIS.2, which is also MIT. Its official pipeline uses the same RMBG-2.0 background remover as TRELLIS.2, so supply a cut-out image if you run the official scripts for paid work.

Which open 3D model works on a smaller graphics card?

By the makers' figures, Hunyuan3D 2.0 needs 6 GB for shape alone and 16 GB with textures, and Stable Fast 3D about 6 GB. TRELLIS.2's card asks for 24 GB, but LocalAI's documentation puts its 1024-resolution pipeline at roughly 10 GB, and Wan2GP ships compressed 8-bit versions.

Which open-source model turns one image into a Gaussian splat?

TripoSplat, from VAST-AI, MIT licensed and built into ComfyUI and Wan2GP. Apple's SHARP also makes a 3D splat scene from one photo in under a second on a standard GPU, by Apple's figures, but its license is research-only.

Why do some open 3D models show zero downloads on Hugging Face?

The Hub counts a download only when certain model files are requested. Pixal3D, TripoSplat, Step1X-3D and UltraShape store their weights in files it does not count, so their official pages record zero even though Pixal3D's ComfyUI copy drew 806,174 downloads in 30 days. App support is a better guide than download counts for 3D models.

This research was produced by Creative AI News.

Subscribe for free to get the weekly digest every Tuesday.