Everyone sees the demos. Type "cyberpunk street food stall", wait 60 seconds, get a 3D model. What the demos don't show is the half of generations that produce mangled geometry, or the export settings that silently break your UVs.
I've been running prompts through text-to-3D tools for the past few weeks — roughly 50 generations of prop assets for a hobby game project. This post is the workflow I settled on, including the steps I skipped at first and regretted.
1. Batch prompts, don't gamble on one
Single-prompt thinking is the #1 workflow mistake. Generation is cheap; your time isn't. For any asset the game actually needs, I run the same prompt through 3-4 generations in one session, then pick based on silhouette alone — texture quality is easy to fix later, geometry isn't.
My prompt template for props:
text
[object] with [2 defining features], [material] [style],
single piece, clean topology, [view]
Example: supply crate with rope handles and stenciled markings, weathered oak, single piece, clean topology, three-quarter view
2. The silhouette test
Before importing anywhere, spin the model in the generator's viewer and judge only the black silhouette. If the outline reads clearly as the object from front, side, and three-quarter angles, proceed. If the back is mangled (extremely common — generators optimize for the view implied by your prompt), regenerate with "full view" or "all sides" added.
This 10-second check has saved me from importing at least a dozen models I'd have spent an hour fixing in Blender before discovering they were unsalvageable.
3. Format choice is a workflow decision
You're targeting
Export
Why
Unity / Unreal
FBX
Animation-friendly, better scale handling
Web / AR preview
GLB
Single file, PBR textures embedded, glTF standard
3D printing
STL
Watertight mesh, no texture baggage
Blender cleanup
OBJ
Lossless geometry, you'll re-UV anyway
One practical note: GLB from most generators embeds textures in a way Unity's importer sometimes doubles. If textures look wrong after import, check the material assignment before blaming the generator — I lost an evening to this.
4. Budget for a cleanup pass — always
Every AI-generated mesh I've shipped needed at least:
- Remesh/re-topo for anything deforming (even slightly)
- Non-manifold repair if it's headed to a physics engine
- LOD generation — generators output one density; your engine wants 3
For background props, a decimate + normal-map bake takes 10 minutes and saves runtime memory. For anything the player touches, plan a real retopo session. There is no prompt that skips this step — yet.
5. The honest scorecard after 50 generations
- ~60% usable for background/props with a cleanup pass
- ~25% garbage — usually over-complex prompts or objects with intricate internal structure (machinery, open interiors)
- ~15% surprisingly great — organic shapes (rocks, plants, food) massively outperform hard-surface tech
Organic > hard-surface is the pattern nobody warns you about. Crates and machinery fight the generator; mushrooms and ruins almost always work. If your game has a forest level, you're in luck. If it's set in a server room, lower your expectations.
What's next
The generators are improving on a ~6-month cadence, and hard-surface quality is the current frontier. My benchmark is simple: I re-run the same supply-crate prompt monthly and compare topology. The day it comes back animation-clean, prop pipelines change forever.
If there's interest, I'll do a follow-up on the cleanup pipeline itself — the Blender side of turning a generated mess into a shippable asset.
Top comments (0)