Every few weeks another AI 3D generator trends — hi3d, Tripo, Meshy, Rodin, Hunyuan3D, and now Microsoft’s TRELLIS.2. The names change faster than anyone can test them.
Rather than another feature table, here are the five questions that actually decide whether a generator is useful to you. Ask them of any tool and you will know within minutes.
1. Does It Output Meshes or Point Clouds?
This is the first filter and it eliminates a surprising number of tools. A point cloud is not an asset — it is data you still have to convert. LingBot-Map is genuinely impressive and outputs point clouds, which makes it useless for someone expecting a model to drop into Unreal.
If the marketing says “3D reconstruction” rather than “3D asset generation,” check this first.
2. Does It Do Real PBR, or Just Colour?
Most generators bake a colour texture and stop. That looks fine in their demo render and wrong the moment you light it yourself, because the surface has no roughness or metallic information.
TRELLIS.2 outputs base colour, roughness, metallic and opacity separately. That is the difference between an asset that behaves like metal and one that is a picture of metal.
3. What Is the Real Hardware Requirement?
“Runs locally” is doing a lot of work in most announcements. TRELLIS.2 officially wants Linux and a 24 GB NVIDIA GPU. Its published “3 seconds” figure is at 512³ on an H100 — 1024³ takes about 17 seconds and 1536³ about 60.
Always ask: at what resolution, on what card. And check whether a community C++ port exists — those have repeatedly dropped requirements far below the official spec.
4. What Does the Licence Actually Say?
“Open source” and “open weights” are different claims, and the gap is where commercial risk lives.
TRELLIS.2 is MIT — but its repository notes that dependencies like nvdiffrast and nvdiffrec carry separate terms. Alibaba’s video ControlNet ships under a community licence with a revenue threshold, not a permissive one. Hosted services often restrict commercial use to paid tiers.
If you are shipping client work, read the actual licence, not the badge.
5. What Is the Topology Like?
Every generator produces topology from an algorithm, not an artist. For props, scans and background objects that is fine. For anything that deforms, it fails the moment you rig it.
This is why retopology tools have been booming alongside AI generation. Generation got cheap; cleanup did not.
The Workflow That Actually Holds Up
- Generate the rough form
- Retopologise — free tools handle props well, characters need manual work
- Re-UV and re-bake, since generated UVs are rarely worth keeping
- Fix the details the model got approximately right
The Honest Summary
AI 3D generation in 2026 is a strong concepting and set-dressing tool with a mandatory cleanup step. The tools improve monthly; the cleanup requirement has not moved.
Anyone selling one-click production assets is describing a demo, not a pipeline. Ask the five questions and you will spot it immediately.
Full tool-by-tool breakdown in our guide to AI 3D model generators.
