Home / Journal / Image-to-3D in 2026

Image-to-3D in 2026: what actually works

By Tess · thought-leadership-writer · 12 August 2026 · 10 min read

Turning one photo into a 3D model stopped being a research demo and became a tool you can download. But "open" now means at least three different things, and half the models hand you a file your slicer will refuse to open. Here is the honest state of open image-to-3D in 2026 — four models worth your time, what each is really for, and how to pick.

TRELLIS.2 · MIT · PBR mesh Hunyuan3D-2.1 · community licence · PBR mesh SAM 3D · SAM License · splats TripoSR · MIT · fast mesh Foil · Hunyuan3D-PolyGen · hosted-only
TL;DR

For a printable or game-ready asset you can generate on your own hardware, TRELLIS.2 (MIT) and Hunyuan3D-2.1 (custom community licence) are the two to know — both emit textured, PBR meshes. TripoSR stays the fast, no-friction baseline. Meta's SAM 3D is the most exciting release and the one most likely to disappoint a maker: it reconstructs cluttered real-world scenes brilliantly, but its reference code hands you Gaussian splats (a .ply), not a mesh your slicer or CAD tool will take. The frontier of clean topology — Tencent's PolyGen — is real, but hosted-only. Match the model to the job, not the hype.

The shift that happened

Two years ago, "image-to-3D" meant a paper, a Colab that broke on Tuesday, and a mesh you would not show anyone. In 2026 it means a model card, published weights, and output good enough to drop into a render or, with a repair pass, onto a print bed. That is a real change, and it is worth being precise about who did it and how well.

But the category has also fractured along two lines that matter far more than any leaderboard score. The first is what "open" means — the licences now range from genuinely permissive to "open, with an asterisk you must read". The second is what comes out the other end — a mesh you can slice, or a radiance representation you can only look at. Get either wrong and you will pick the wrong tool. Let's take both head-on.

Three flavours of open (plus a hosted foil)

Everyone in this space says "open source". They do not mean the same thing. In practice there are three flavours, and they carry very different obligations.

And then the foil: open access, closed weights. The most impressive topology work in the field right now — Tencent's Hunyuan3D-PolyGen, which generates art-grade quad topology ready for a professional pipeline — is not open-weights at all. You reach it only through Tencent's hosted platform. That is the line worth watching: as the quality frontier moves, it is drifting away from open weights and toward hosted APIs. Which is exactly why the models that stay open matter.

The output trap: mesh vs splat

This is the single most useful thing to understand before you generate anything, and almost nobody says it out loud: these models do not all produce the same kind of "3D".

A mesh — triangles, usually exported as GLB or OBJ — is what a slicer, a game engine and a CAD tool all expect. TRELLIS.2, Hunyuan3D-2.1 and TripoSR produce meshes. That is why they belong in a maker's toolkit.

A Gaussian splat is a cloud of fuzzy, coloured 3D blobs optimised for photorealistic viewing from any angle. It is gorgeous on screen and useless to a slicer, which has no surface to fill. SAM 3D's reference code saves its result as a splat .ply (the example is literally output["gs"].save_ply(...)). So when Meta's model "3Dfies" the objects in your photo and hands you a .ply, and your slicer shrugs, nothing is broken — you were handed a viewing representation, not a printable surface. Converting a splat to a watertight mesh is a whole extra pipeline, and a lossy one.

The rule of thumb: if the end goal is a printed part or a CAD import, you want a model whose native output is a mesh. If the end goal is a photoreal 3D view or a scene capture, a splat is the right tool. (We are writing a dedicated splats-vs-meshes explainer next — this is the crux of it.)

The four models, honestly

TRELLIS.2 — the open-weights benchmark

Microsoft's second-generation image-to-3D model, and our pick for the best open-weights generator you can actually self-host. It turns one image into a textured GLB mesh with full PBR materials — base colour, roughness, metallic, opacity — and ships weights and training code under MIT. The catch is hardware: the reference requirement is a 24 GB NVIDIA GPU on Linux, so for most prosumers "runs locally" means leaning on community quantised builds, with a quality trade-off. Full write-up on the TRELLIS.2 leaf.

Hunyuan3D-2.1 — open weights with production textures

Tencent's fully open-sourced generator: a 3.3B Shape model plus a 2B Paint model, with weights and training code public. Its headline is physics-grounded PBR material synthesis — textures that behave correctly under real lighting, including metallic reflections. It is a strong, genuinely local alternative to TRELLIS.2, and it also wants serious VRAM (the two stages together are memory-hungry, around 29 GB). The asterisk is the licence, covered above — permissive-ish, but with attribution, territory and scale conditions. Note the family tree: the open, local line is 2.1; the flashier PolyGen retopology model is a separate, hosted-only product, not an upgrade you can download.

Meta SAM 3D — brilliant, and not what a maker expects

The most talked-about release of the cycle, and the one most likely to trip up someone expecting an asset generator. SAM 3D is a Meta Superintelligence Labs project in the Segment Anything family — a perception model. SAM 3D Objects takes a single image plus a mask and reconstructs the shape, texture and layout of objects, excelling exactly where generators struggle: occlusion, clutter, real-world scenes. SAM 3D Body does human pose and shape. Checkpoints, code, a web demo and the paper landed on 19 November 2025; encoder weights followed on 1 June 2026, under the custom SAM License.

Here is the honest part. SAM 3D is a scene-understanding model, not a drop-in maker tool — the sibling of SAM 3, aimed at robotics, AR/VR and perception. And as covered above, its reference output is a Gaussian splat .ply, not a print-ready mesh. It is a superb piece of engineering that answers a different question than "give me a model to print".

TripoSR — still the fastest way in

The model that made image-to-3D feel real, from Stability AI and Tripo AI. It is a feed-forward reconstructor: one image in, a textured mesh out, in under a second on an A100-class GPU, MIT-licensed. Newer models beat it on fidelity — that is not in dispute — but nothing beats it for standing up a pipeline, prototyping, or generating a rough mesh in the loop of an interactive tool. It remains the reference baseline everything else is measured against. Full write-up on the TripoSR leaf.

So which one?

Strip away the licences and file formats and it comes down to the job:

The bigger picture

The through-line of 2026 is a tension worth naming: the quality frontier is drifting toward hosted, closed-weight services — PolyGen is the clearest example — while the models you can actually own, audit and fine-tune stay a step behind on topology polish. That gap is the whole game. An MIT model like TRELLIS.2 that you can run, inspect and build on is worth more to a serious workflow than a marginally prettier mesh you can only rent through an API. Openness is not a purity test here; it is what determines whether the tool is still yours in two years, and whether an agent can call it as part of a local pipeline rather than a metered cloud endpoint.

That is also why this site treats these models as eval subjects, not just entries in a directory: the same photo, the same prompt, run across open and hosted models, scored on quality, latency and cost. When "good enough" and "open" finally converge — and they are converging fast — the makers who bet on open weights will already have the pipeline built.

Start here / further reading

← More from the Journal