Picture a three-person indie team building a fantasy tavern level. They need fifty-odd props: barrels, lanterns, tankards, a moth-eaten stag head over the fireplace. A few years ago, every one of those meant concept art first, then hours in Blender. Today they can sketch a direction in the morning and have rough 3D drafts on screen by lunch.
That speed is real. It’s also where many teams get into trouble, because a draft that looks finished usually isn’t.
Going from concept image to production asset has always involved a chain of specialist steps: visual development, modelling, topology, UVs, texturing, rigging, optimisation and engine integration. Tools like Meshy’s AI art generator for game development, which feeds directly into image-to-3D conversion, change the speed and order of that chain. They don’t remove any links from it.
Where the industry actually stands
Developers aren’t of one mind about this. In GDC’s 2026 State of the Game Industry survey of more than 2,300 professionals, 36% said they use generative AI tools in their work. Yet 52% said generative AI is hurting the industry, up from 30% the previous year. Artists were the most sceptical: 64% of people in visual and technical art roles took the negative view.
That split is worth keeping in mind. The teams getting value from AI tend to use it to speed up the dull parts, not to replace the craft.
Start with a picture worth building
Everything downstream depends on the reference image. A good one has a readable silhouette, materials that make sense, and enough detail to explain what the object is for. Our tavern team’s lantern needs to read as “iron frame, glass panes, candle inside” at a glance, or the 3D step will guess wrong.
Image generators are useful here as an exploration tool. You can try a dozen lantern styles, three levels of wear and a couple of colour palettes before committing. Meshy’s generator, for instance, accepts up to five reference images, offers style presets from anime to Western realism, and returns four variations per prompt. It also has pose and multi-view settings meant to produce images that convert to 3D more cleanly.
Once a direction is picked, an artist should review it before moving forward. Is the concept original enough? Can you clearly understand the design from every angle? A pretty image with a confusing back side can easily turn into a confusing 3D model.
The first 3D draft: fast, and full of guesses
Image-to-3D systems infer a mesh and textures from what they can see. For props, background objects, and early prototypes, that’s dramatically faster than blocking out shapes from an empty scene. The team gets something they can spin around, compare, and drop into a test level.
The word to remember is infer. The system has to invent everything the camera didn’t show: the back of the barrel, the underside of the table. It often misreads shadows as shapes, struggles with hair and glass, and merges overlapping parts.
So judge the draft from every angle, not just the one that matches the reference. The front usually looks great. The back is where the surprises are.
Geometry: what’s good enough depends on where it’s going
How much clean-up a model needs depends entirely on its destination.
A still render for a pitch deck can get away with messy geometry. A game asset can’t. It needs a sensible polygon budget, shading that doesn’t flicker, clean UVs, and topology that bends properly if the object moves.
AI tools help a bit. Meshy’s remesh feature, for example, lets you target a specific polygon count. But artists will still retopologise key areas, delete hidden or floating geometry, fix proportions, and rebuild thin details like rope, handles, and fingers. Characters need the most care around faces, hands, clothing and joints, because that’s where bad topology shows the moment they move.
Textures: watch for baked-in lighting
Generated textures communicate the look quickly. They also tend to carry problems: shadows and highlights painted straight into the base colour, uneven detail, and smeared nonsense on surfaces the AI never saw.
A production workflow pulls those apart. Many generators can output PBR maps (base colour, roughness, metallic, and normal maps, the standard set game engines use), but each one still needs checking against your renderer. Look at seams, texture resolution, and whether the metal parts actually behave like metal.
Consistency across a set matters as much as quality. One beautifully detailed tankard looks wrong next to forty flat, stylised ones. The same principle that keeps characters consistent across AI image batches applies here: fix the shared rules first, then generate within them.
Rigging, optimisation and the engine test
Anything that moves needs a skeleton, skin weights, test animations and collision. Some tools now offer automatic rigging. Meshy, for one, supports auto-rigging for humanoid and quadruped models, which is a big time-saver for background creatures and NPCs. Expect to fix weights by hand around shoulders, hips, and anywhere clothing overlaps.
Static props have their own checklist: correct real-world scale, sensible pivot points, clear naming, and levels of detail for distance. Web and mobile projects add strict limits on file size and texture memory on top.
The single most useful habit is to get the asset into your actual engine early. A model that looks perfect in a generator’s preview can show flipped normals, wrong orientation, odd shading or a frame-rate hit once it’s in Unity or Unreal under real lighting. Test in context before polishing anything.
Don’t skip the rights check
Teams often forget this step until it’s expensive.
Check the licence terms of every tool in your pipeline. Meshy, for example, grants full commercial usage rights to paid-plan users, while free-tier outputs are for personal and non-commercial use only. A prop made on a free account and shipped in a commercial game is a problem waiting to happen.
Platforms have rules too. Steam has required developers since 2024 to disclose AI-generated content used in their games. Plan that disclosure from the start rather than scrambling at submission.
Set quality gates, not just deadlines
The teams that do this well don’t let assets slide through the pipeline on vibes. They define gates.
| Gate | Who signs off | What they check |
| Concept | Art lead | Originality, clarity, style fit |
| Mesh | 3D artist | Silhouette from all angles, polycount, topology |
| Materials | Texture/tech artist | PBR maps, baked lighting removed, set consistency |
| Rig | Animator | Deformation, weights, test animations |
| Performance | Tech artist | Engine test, LODs, memory budget |
| Rights | Producer | Tool licences, disclosure requirements |
Our tavern team might push forty background props through these gates in a week. The hero items, like the fireplace stag head the player stares at for a cutscene, still get a full manual pass.
The short version
Generative AI shrinks the distance between “here’s an idea” and “here’s a model we can test.” That frees artists to spend more time choosing, refining and telling the story, and it makes trying a wild idea cheap.
“Production-ready” is still a judgment call, not a button. AI speeds up the handoffs between stages. Experienced people still decide when an asset is actually done. For a broader look at the tools themselves, here’s how AI 3D model generators compare as a starting point for creators.
Related: What Does “Timing” Actually Mean in Games? Frames, Lag, AI & More
