Hunyuan3D 2.1: the open-weight 3D generation baseline and default starting point for self-hosting
hunyuan3d-2
Tencent released weights for both the shape and texture stages plus PBR material output, which gives 3D generation its first reference implementation you can run on your own machine, fine-tune, and wire into an internal pipeline. Its value is not a leading score but reproducibility: any claim about 3D generation quality can now be measured against it.
- CONFIDENCE
- Vendor Claim
- Official model card or keynote only, no independent re-test
- MATURITY
- Product
- research → demo → product → production
Our take<p>We put this on the 3D & world-model ladder for a completely different reason than Tripo: Tripo tops the ladder on <strong>delivery capability</strong>, Hunyuan3D 2.1 tops it on <strong>verifiability</strong>. Open weights mean its failure modes can be reproduced and argued about by anyone, which is rare in generative 3D.</p><p>To be fair about the cost: both the shape and texture stages want serious VRAM, and running batches in production means building your own queueing, caching and retry logic, because the reference inference scripts are research-oriented rather than service-oriented. The PBR material output is first-tier among open options but still short of commercial stability. If the goal is "a pile of assets today", call a commercial API. If the goal is "build real judgement about 3D generation, or fine-tune for a domain", this is the best starting point available right now.</p>
Two stages: why shape and texture are separate
Hunyuan3D 2.x keeps the "shape first, texture second" split: the shape stage is a diffusion transformer that produces geometry exportable as a mesh, and the texture stage synthesises maps and PBR channels on top of it. The practical benefit is that you can replace half of it - if you already have good geometry (scan, CAD, sculpt) you can run only the texture stage, and when fine-tuning for a domain you can touch the shape stage without disturbing textures.
The 2.1 line makes PBR output and higher-resolution textures the headline upgrades, which is what first lets its output enter engines that assume a metallic/roughness workflow rather than serving only as a preview model.
The honest self-hosting checklist
The repository ships research scripts. Turning that into a service means filling in a fair amount:
- VRAM and concurrency: single-GPU serial inference is the default posture; batches need a queue, and the texture stage peaks higher than the shape stage.
- Output governance: face count, UV island count and texture size are unconstrained, so a normalisation pass (decimation, re-unwrap, compression) belongs in front of your pipeline.
- Retries: some prompts produce degenerate geometry (cavities, shards, non-manifold edges); detect and regenerate automatically.
- Licensing: Tencent's open weights come with their own licence terms. Read them before commercial use - "open source" does not mean "unrestricted".
Boundaries and failure modes
- Assemblies and multi-object scenes: like every current 3D generator, it is good at one subject and unreliable on relations and occlusion between several.
- Thin and openwork structure: tends to fill in or break.
- Texture/geometry misalignment: back-facing texture is inferred, and single-image input produces visible seams.
- Scale: no real units; calibrate by hand.
- Long-tail categories: training is weighted toward common objects, so specialist equipment, instrument internals and mechanical parts return "resembles but is not".
How to use it well
Three postures we recommend. First, as a baseline: when evaluating any commercial 3D service, run the same prompt set through Hunyuan3D - the delta is what you are actually paying for. Second, as a fine-tuning base: for a narrow category (footwear, bags, furniture, toys) with a few hundred to a few thousand in-house assets, fine-tuning the shape stage pays off quickly. Third, as an offline rough-model machine: batch overnight, filter by hand in the morning, and never expose the variance directly to users.