Every few weeks a new AI 3D generator trends — hi3d, Hyper3D, Tripo, Meshy, Rodin, TRELLIS.2. The names change faster than anyone can evaluate them, and most coverage just repeats the launch post.
Rather than another leaderboard, here is what has genuinely changed in the last year, and what has not.
What Changed: Materials
A year ago, image-to-3D meant a mesh with a colour texture. That looks acceptable in the demo render and wrong the moment you light it yourself, because the surface carries no roughness or metallic information.
TRELLIS.2 outputs base colour, roughness, metallic and opacity as separate properties. That is a real step — the difference between an asset that behaves like metal and a picture of metal.
What Changed: Hardware Access
Models ship needing datacentre GPUs, and the community ports them down. TRELLIS.2 officially wants Linux and 24 GB of VRAM; a C++/GGML port runs the same pipeline without CUDA. The same happened to NVIDIA’s text-to-motion model with kimodo.cpp.
That pattern is now reliable enough to plan around: if a model matters, expect a port within weeks.
What Has Not Changed: Topology
This is the constant, and it is why the “production-ready” claims keep failing.
Every generator produces topology from an algorithm. For props, scans and background objects that is fine. For anything that deforms, it fails the moment it is rigged — loops are not where a face or joint needs them.
Generation got cheap. Cleanup did not. That is why retopology tools have been growing alongside AI generation rather than being replaced by it.
What Has Not Changed: Licensing Confusion
“Open source” and “open weights” are still used interchangeably, and they are not the same. TRELLIS.2 is MIT but its dependencies carry separate terms. Community-licensed models often have revenue thresholds. Hosted services frequently restrict commercial use to paid tiers.
Read the actual licence before shipping client work.
The Realistic Position
AI 3D generation in 2026 is a strong concepting and set-dressing tool with a mandatory cleanup step. That has been true for a while and remains true despite each new release being announced as the one that changes it.
Our framework for evaluating any of them is in five questions that matter.
