Hunyuan3D 2.1 vs TRELLIS.2 vs TripoSR (2026): Which Open Image-to-3D Model Wins?

Nobody here ships an independent, apples-to-apples benchmark you can just trust, so treat every "best quality" line as a starting point, not proof. What you can compare is how each model builds its shape. TRELLIS.2 replaced the original TRELLIS's Structured Latent pipeline with O-Voxel, a "field-free" sparse voxel representation that Microsoft designed specifically to escape the limits of iso-surface fields such as SDFs and Flexicubes. The practical claims are concrete: open surfaces like clothing and leaves, non-manifold geometry, sharp features, and fully enclosed internal structures, generated at 512-cubed, 1024-cubed, or 1536-cubed voxel resolution. That is a different class of geometry from the original TRELLIS, whose own project page conceded limited capability on photorealistic real-world objects.

Hunyuan3D 2.1 is the proven alternative: its diffusion-transformer shape stage produces meshes that Tencent's Hunyuan3D 2.0 report describes as holeless, the property that makes a mesh usable downstream without a repair pass, and it has a year of production use behind it. TRELLIS.2 is candid about the other side of its topology freedom: Microsoft's model card notes that raw meshes may occasionally contain small holes or minor topological discontinuities, ships a hole-filling step for strictly watertight needs such as 3D printing, and describes the 4B checkpoint as a base model not yet aligned to human preferences. TripoSR is the honest weak link on geometry: it reconstructs from a single view in one feed-forward pass, so it is fast and clean but carries less fine detail than a multi-step model, and it has no way to reason about the parts of the object the photo does not show. On the documentation, TRELLIS.2 takes this round for resolution and topology range, with Hunyuan3D's holeless meshes the safer, better-tested bet if your pipeline needs watertight output without a post-process.

This used to be Hunyuan3D's clearest win, and it is now a genuine contest. Hunyuan3D 2.1 runs a dedicated second stage, Hunyuan3D-Paint, that synthesizes real PBR materials onto the mesh: albedo, metallic, and roughness maps rather than baked-on vertex color, painting from up to nine rendered views at up to 768 resolution and baking 4096-pixel texture maps. That is exactly what a game or product-visualization pipeline wants, and it has been shipping since June 2025.

TRELLIS.2 changes the picture because materials are native to its representation rather than a separate pass. O-Voxel stores base color, roughness, metallic, and opacity alongside the geometry, so every generation comes out as a PBR-ready GLB, and the opacity channel means translucent surfaces such as glass or foliage are modeled rather than faked. Texture resolution is not the differentiator: both bake 4096-pixel maps in their reference pipelines. TRELLIS.2 also has a shape-conditioned texturing mode for meshes you already have, matching Hunyuan3D's texture-only mode. On the documentation, TRELLIS.2 now edges this round on completeness, chiefly the opacity channel, with two caveats: this is a read of official material, not a side-by-side render test, and the GLB ships in opaque mode by default so you enable transparency in your DCC tool. TripoSR is the honest weak link: it outputs vertex colors only, with no PBR materials and no separate texture-synthesis stage, and Stability itself now points users toward its faster, game-ready successor SF3D for better surfacing.

Here the generational change cuts the other way. The original TRELLIS was the most flexible tool in this category: one Structured Latent result decoded to a textured mesh, 3D Gaussians, or a radiance field, and it shipped both image-to-3D and text-to-3D models. TRELLIS.2 narrows that on purpose. It is a single-image, image-to-3D model whose output is one thing, a mesh with PBR materials as GLB, and there is no text-conditioned TRELLIS.2 checkpoint. If you need Gaussians, a NeRF, or a text prompt from Microsoft's stack, you still run the original TRELLIS, and that is fine, the repo is still up and MIT, but it is a separate, older model.

Hunyuan3D 2.1 is therefore the broader open workflow today: it exports GLB and OBJ, accepts a single image or multi-view input through the 2mv variant in its 2.0 family, and has a texture-only mode that paints materials onto a mesh you already have. Its open weights are image-conditioned: the repo's text-to-3D mode works by chaining a text-to-image model (HunyuanDiT) in front of the image-to-3D stage, so there are no text-conditioned 3D weights. TripoSR is the narrowest of the three on purpose: a single image in, one mesh out as OBJ or GLB, with an optional baked-texture UV. No text input, no multi-view, no alternative representations. That simplicity is the trade for its speed and small footprint. For input options in a single open release, Hunyuan3D 2.1 takes this round.

Speed is where TripoSR still pulls ahead, and it is a structural advantage rather than a tuning trick. TripoSR is a feed-forward reconstruction model: one pass through the network turns a single image into a mesh, and the paper and repo both put that at under 0.5 seconds on an NVIDIA A100. There is no diffusion or flow sampling loop to wait on. For a generate, glance, iterate loop it is a different experience from waiting on a multi-step model.

The gap has narrowed, though, and TRELLIS.2 deserves credit for publishing real numbers where most projects do not. Microsoft reports total shape-plus-material time of about 3 seconds at 512-cubed resolution, about 17 seconds at 1024-cubed, and about 60 seconds at 1536-cubed, measured on an H100. So at its lowest resolution TRELLIS.2 is only a few seconds behind TripoSR while producing a full PBR asset, and the price of its top resolution is a minute per object. Hunyuan3D 2.1 publishes no per-asset latency, so we will not invent one; its two-stage shape-then-paint pipeline is inherently heavier than a single pass, which is why distilled Turbo and Fast variants exist in the 2.0 family. The honest caveat on TripoSR's number: that sub-second figure is raw model inference on a top-tier GPU, so add background removal and slower hardware and your real end-to-end time grows. Relative to the other two, TripoSR is still the fast one.

This is where the choice often gets made for you, because it decides what you can actually run. TripoSR needs about 6 GB of VRAM for a single image and falls back to CPU when no CUDA device is present, so it runs on modest, older, or laptop-class cards. That is a genuine advantage: it is the one most people can run on hardware they already own.

The two heavyweights ask for more, and TRELLIS.2 asks for more than its predecessor did. Microsoft states an NVIDIA GPU with at least 24 GB of memory, up from the original TRELLIS's 16 GB, verified on A100 and H100, and the code is tested only on Linux. That rules out every 16 GB consumer card and most 12 GB ones for a stock install. Hunyuan3D 2.1 needs 10 GB for shape generation alone, 21 GB for texturing, and 29 GB for the full shape-and-texture pipeline, which puts a complete textured result out of reach on most consumer cards, though a low-VRAM mode and the lighter 2mini variant help if you have less memory, and the older Hunyuan3D 2.0 pipeline fits a full textured result in about 16 GB (6 GB for shape alone) if you can live without 2.1's PBR paint stage. So the practical ranking is clean: TripoSR runs almost anywhere, TRELLIS.2 wants a 24 GB Linux box, and Hunyuan3D wants a bigger one for its best output. If your GPU is the constraint, this dimension can settle the whole comparison before the others matter.

The best model is the one you can actually get running and keep using, and here Hunyuan3D still has the most momentum in the community even though its open repos have been quiet since late 2025. The flagship Hunyuan3D-2 repo carries the biggest star count of the three, the family ships several variants (mini, multi-view 2mv, distilled Turbo and Fast), community ComfyUI and Unity integrations are linked from the official README, the 2.1 README lists macOS, Windows, and Linux support, and Tencent runs a hosted platform if you do not want to run it locally at all. That breadth is a real advantage when you hit an edge case and need an alternative variant or a community node. Worth knowing: Tencent's newer open 3D work has gone into sibling projects (Hunyuan3D-Omni, Hunyuan3D-Part, world models) rather than a 2.1 successor with open weights.

TRELLIS.2 is the fast-rising one: the repo passed 11,000 stars in nine months, it ships an official Hugging Face Space demo, a Gradio app, a texturing app, and full training code, and Microsoft was still pushing commits in mid 2026. Since ComfyUI v0.34.0 (August 2026) it also runs natively in ComfyUI core with official weights and no custom nodes. What it still lacks is a hosted product and any official non-Linux support, and its reference setup script compiles several CUDA extensions. TripoSR is the simplest to get running (a single image, about 6 GB of VRAM, a ready Hugging Face Space, and a Gradio app) but it is the oldest and smallest ecosystem here, its ComfyUI support is community-provided rather than official, and Stability now steers users toward SF3D. For the richest tooling, Hunyuan3D is the pick; for the shortest path from zero to a mesh, TripoSR is.

All three are free to download, but "open" splits into details that matter the moment you want to ship something commercial. TRELLIS.2 is MIT for the model and the code, with one explicit caveat in its own README: two NVIDIA dependencies it uses for rendering, nvdiffrast and nvdiffrec, carry their own separate license terms, so a fully clean commercial stack takes a quick dependency check (the original TRELLIS had the same shape of caveat). TripoSR is plain MIT: use it, modify it, sell what you make, no user cap, no regional carve-out, no strings.

Hunyuan3D is the outlier, and it is worth being clear-eyed about. Its weights ship under the Tencent Hunyuan 3D 2.1 Community License, which is not MIT. Two clauses stand out: if your product crosses 1 million monthly active users you must request a separate license that Tencent may or may not grant, and the license territory explicitly excludes the European Union, the United Kingdom, and South Korea. It also forbids using outputs to train non-Hunyuan models. So Hunyuan3D is free to download but not unrestricted, and for an EU or UK builder that regional exclusion can be a hard stop no matter how good the model is. If "open and no strings" is what you need, TripoSR and TRELLIS.2 are the cleaner answers.