Category 1: prompt-to-asset assistants (text to 3D in Blender)
What it does: turns a written brief into geometry inside your open scene - props, furniture, buildings, interiors, blocked-out characters.
Where it wins: blocking out, kitbashing, set dressing, and producing variants of an asset fast enough to review with a client on the call.
Where it stops: hero characters with heavy deformation, anything that must match an existing art direction exactly, and real-time budgets that need a deliberate optimisation pass.
The tool to look at: Blender AI. It runs in-scene, produces quad-dominant geometry at real-world scale, and keeps the result as ordinary Blender data you can retopologise and export. Plans start at $9 a month for 100 prompts, and there is no free tier.
Category 2: AI texture generation
What it does: generates or edits textures and materials - tileable sets, damage passes, stylised surfaces - from a prompt or a reference image.
Where it wins: environment work, mood passes, and anything where you need twenty plausible surfaces instead of two perfect ones.
Where it stops: precise brand colours, technical PBR values that have to match a supplier's data, and anything that must be identical across a set of assets.
The tool to look at: Dream Textures remains the reference open-source option for Blender, and it runs locally, which matters when the material is client-confidential.
Category 3: AI-assisted modelling and procedural work
What it does: combines Geometry Nodes with model-driven helpers - generating variation across a set, adapting a kit to new dimensions, or driving repetitive node graphs from a description.
Where it wins: modular kits, scattering and population, and parametric families where the same asset needs forty versions.
Where it stops: it is not a modelling shortcut. You still author the node graph; AI helps you build, vary and maintain it.
The reason this category matters for AI 3D work is scale: a generated asset that becomes a node graph variant is reusable, while a generated asset that stays a one-off is not.
Category 4: scan and photo cleanup
What it does: photogrammetry with AI cleanup - turning photos into meshes, then repairing the usual scan problems: noise, holes, non-manifold edges, absurd polycounts.
Where it wins: real-world objects and locations that would take days to model: a specific chair, a heritage facade, a product with detailed moulding.
Where it stops: it reproduces what exists. For anything that has to be changed, you are back to modelling.
The tool to look at: Poly.cam for capture, then Blender for the cleanup pass - merging vertices, rebuilding flat surfaces, decimating to a budget.
How to choose, in one table
| Your job | Category that pays off | Practical first pick |
|---|---|---|
| Block out props and environments fast | Prompt-to-asset assistant | Blender AI |
| Produce dozens of surfaces quickly | AI texture generation | Dream Textures |
| Maintain parametric asset families | AI-assisted procedural modelling | Geometry Nodes + AI layer |
| Digitise real objects or places | Scan and photo cleanup | Poly.cam + Blender cleanup |
What to check before you commit to any tool
- Where the output lives. In-scene generation keeps units and context; file-based generation adds an import and repair step.
- What the licence says. Commercial rights differ between tools, and client work is where that becomes expensive.
- How it is metered. Prompts, credits or monthly caps - pick the model you can budget for rather than the one with the best demo.
- Whether it survives export. A generator that produces clean Blender geometry is always more useful than one that produces a screenshot.