Search interest in AI 3D generators spiked again this week, with TRELLIS.2 and a cluster of newer names breaking out. The pattern is now familiar enough to be worth naming, because it repeats every few weeks and it costs people time.
The Cycle
A model is announced with striking demo footage. Search volume spikes within days. Roundup posts appear, mostly rewriting the announcement. Three weeks later the search volume is gone and almost nobody who tried it put the output in a project.
That is not because the models are fake. It is because generating a mesh and using one are different problems, and the gap between them has not moved much.
What Has Genuinely Improved
Materials. A year ago image-to-3D meant a mesh with a colour texture baked with lighting in it. TRELLIS.2 outputs base colour, roughness, metallic and opacity as separate properties. That is the difference between an asset that behaves like metal and a picture of metal.
Hardware access. Models ship needing datacentre GPUs and the community ports them down within weeks — trellis2.cpp runs the same pipeline without CUDA. That pattern is reliable enough to plan around.
What Has Not
Topology. Every generator produces geometry from an algorithm, not from an artist. For props, scans and background objects that is fine. For anything that deforms it fails at rigging, because loops are not where a joint needs them.
Generation got cheap. Cleanup did not. That is why retopology tools have grown alongside AI generation rather than being replaced by it.
Licensing clarity. “Open source” and “open weights” still get used interchangeably and are not the same. Dependencies carry their own terms, community licences often have revenue thresholds, and hosted services frequently restrict commercial use to paid tiers.
How to Evaluate One in Ten Minutes
Ignore the hero render. Ask five things:
- What does it output? Mesh, point cloud or radiance field — these are not interchangeable.
- Does it produce PBR maps, or one texture with lighting baked in?
- What hardware, honestly — and does it run on Windows or Linux only?
- What is the actual licence on the weights and the dependencies?
- Will it deform? If the answer is no, it is a prop tool.
Our longer version is in five questions that matter.
The Realistic Position
AI 3D generation in late 2026 is a strong concepting and set-dressing tool with a mandatory cleanup step. That has been true for a while and remains true after each release that is announced as the one that changes it.
Which is fine. Concepting and set dressing are real work, and doing them faster is worth having. It is just not the claim being made.
