Generate 3D Asset
Generate a 3D model from a text prompt or reference image. Drag the result into your scene like any library asset.
Type a description, drop in a reference image, or hand the system three photographs from different angles. The output is a 3D mesh ready to drag into your scene like any other library asset.
Watch
What it does
Three input modes, all producing a usable 3D mesh:
Single reference image. Upload one photograph; the system generates a mesh that matches the visual.
Multi-view reference. Upload up to four photographs of the same object from different angles (front, back, left, right). The mesh honors the geometry implied by the multiple angles much more accurately than single-image. Best for objects with complex geometry, holes, or distinct features on multiple sides.
Text only. Describe what you want in plain language. "A medieval gauntlet, articulated, dark steel." The system invents the geometry from the description.
The result is a mesh that lives in the asset library next to your imports and library items. Drag, swap, all the same workflows as a regular asset.
How to use it
Open the asset library. Bottom of Build mode.
Click Generate 3D (next to Import). The Generate-3D modal opens.
Pick the input mode. Single image, multi-view (front/left/back/right), or text.
Provide the input. Upload the image(s) or type the description.
Set the Detail tier. Standard, Medium, or High. Medium is sufficient for most objects, especially if you're attaching image reference at render time anyway. High is worth the extra cost for objects with complex geometry – lots of holes, fine structural detail, or many distinct surfaces (a bicycle, a piece of machinery). High also performs better when used with multiple image references.
Leave "Check to review and edit" on. The generator doesn't know the real-world scale of the object – results often come in significantly too large or too small. The review step is where you catch and fix that before the mesh lands in your library. Leave it on.
Click Generate. The job runs in the background. A toast appears with progress; the asset library reopens when the result arrives.
Review the result (if the toggle was on). The same editing environment as Smart Import: set scale, orientation, name, description.
Save. The mesh lands in your asset library, available to drag into the scene like any other asset.
When to reach for which input mode
Single image – fastest, cheapest, lowest fidelity. Good for "approximately right" mass that you'll attach an image reference to anyway.
Multi-view – much better geometry, at a higher generation cost (costs are billed in credits – see Credits and tokens). Use when the object has distinct front/side/back features (vehicles, characters, hardware with handles or controls).
Text only – when there's no reference image at all, or when you're brainstorming and the prompt itself is the test ("a robot vacuum that looks like a manta ray").
Custom assets at scale
Once a generated asset is in your library, it behaves like any other library asset. The team-asset visibility pattern works for the agency case where one creative generates a fleet of vehicles for a campaign and the rest of the team can drag them in without re-generating.
The webinar's automotive walkthrough shows the canonical flow:
Generate the hero asset from multi-view reference photographs.
Save to Team Assets with a descriptive name.
Every team member can drop the same hero into different shots and scenes.
Image reference on the hero locks the visual look at render time.
Limits and known issues
Single-image generation can produce ambiguous geometry. A photo of a car from the front gives the system no information about the back. The result will invent something that may not match the actual rear of the vehicle. Use multi-view when accurate geometry matters.
Topology is generation-time, not production-grade. Generated meshes have non-uniform topology that's fine for scene blocking and rendering through the visualizer, but not for downstream pipelines that need clean retopology (Unreal cinematics, real-time engines). For those, generate as starting state and retopologize externally.
No materials. Generated meshes are imported without textures – materials are stripped on the way in. The visualizer reads the object's attached image reference at render time instead. If the images used to generate the mesh were uploaded as reference, they're already attached automatically (see above). For brand-accurate surface treatment, add or replace image reference on the object after import.
Completed generations cost credits even when the result is unusable. Failed attempts aren't charged (billing mechanics live in Credits and tokens). Leave the "Check to review and edit" toggle on so you can inspect before the mesh lands in your library.
Related
Last updated
Was this helpful?