Guides · 3 min read

Using AI image-to-3D in a real game project

What generated meshes are actually good for in a game, what they are not, and how to get one from a photograph into Unity or Unreal with a rig on it.

20 August 2026

There is a gap between "this tool made a 3D model from a photo" and "this asset is in my game." Most of what is written about image-to-3D stops at the first one. This is about the second.

What generated meshes are good for

Props. Overwhelmingly, props.

A generated mesh is strongest on a single object with a clear silhouette: a crate, a lantern, a helmet, a chair, a weapon. These are also the assets that eat the most time in a small project, because each one is a day of modelling for something the player walks past.

They are weakest on three things, and it is worth knowing which before you spend an afternoon fighting one:

  • Scenes with several objects. The model has to guess which thing you meant. It usually guesses the largest, and merges the rest into it.
  • Reflective and transparent surfaces. Glass and chrome give the reconstruction nothing to work from, because what it sees is the room, not the object.
  • Anything whose back is not implied by its front. One photograph contains no information about the far side. The model invents something plausible. For a crate that is fine; for a character's face it is not.

The topology question

Generated meshes are triangulated and dense. For a prop the player never deforms, that is not a problem. It is what a decimator is for, and both engines handle it.

There is a step-by-step for that side of it in getting a generated model into Unity and into Blender.

It becomes a problem the moment something has to bend. A triangulated mesh with uneven density does not deform cleanly at a joint, and no amount of weight painting fixes topology. If a model is going to be animated by hand or subdivided, remesh it to quads first. That is a one-click step here and it costs three tokens; doing it afterwards in a DCC tool costs an afternoon.

Getting it into an engine with a rig

The step most people miss is that a generated character can be rigged without leaving the browser, and that changes the maths of whether this is worth doing.

  1. Generate the character from a photograph or a prompt. If you are prompting, turn on rig-ready. It rewrites the prompt to force a neutral T-pose, which is what automatic rigging needs. It costs nothing.
  2. Rig it. A standard humanoid skeleton is fitted automatically. Non-humanoid meshes are refused rather than rigged badly, which is the right answer.
  3. Apply a motion, or skip that and bring your own clips.
  4. Export FBX. GLB is the better format for the web, but FBX is what carries a skeleton reliably into Unity and Unreal.

From there it is an ordinary imported character. The rig is a standard skeleton, so a humanoid retargeting setup you already have will drive it.

What to check before you commit

Scale. Generated meshes come back in arbitrary units. Set your scale once on import rather than per asset.

Origin. The pivot is usually at the mesh centre, not at the feet or the base. For anything that gets placed on a floor, move it once in your DCC tool or your engine's import settings.

Texture resolution. The standard quality tier returns textures sized for a prop seen at middle distance. If the asset is a hero object the camera gets close to, generate it at high or ultra instead. That is what the multiplier buys, and regenerating later costs more than getting it right first.

Watertightness, if you are printing. Generated meshes are built to look right, not to be manifold. A slicer will tell you.

The honest cost comparison

A prop from a photograph is one token on the Fast engine, four on Studio. Rigged and animated, a character is around eleven. Against a day of modelling time, that is not a close comparison for set dressing.

It is a much closer comparison for hero assets, and for those the answer is usually still a modeller. What this replaces is the twenty things nobody wanted to model, not the one thing somebody wanted to.

Where this goes wrong

The failure mode worth naming: generating fifty props before checking that one of them works end to end in your engine. Do the whole pipeline once on a single asset: generate, remesh, rig, export, import, place, light. Everything you learn from that first one applies to the other forty-nine, and the ones you would have had to regenerate are still unmade.

Share this X LinkedIn
Read next
Turn an image into a 3D model