Modelling without reference produces the version of an object that lives in your head, which is usually a simplified average of everything you have seen.
What reference is for
- Proportion. The most common failure in modelled objects, and the hardest to spot from memory.
- Detail scale. How big is a bolt relative to a panel, a handle relative to a door.
- Construction logic. How the parts actually go together, which is what makes an object read as real.
- Wear story. Where an object gets damaged and why. Uniform wear reads as fake.
Gathering it
- Multiple angles, always. A single front view leaves the depth to invention.
- Detail shots of the parts you will model closely.
- Scale references — something of known size in frame.
- Similar objects, not just the exact one. Understanding the category tells you what is variable and what is fixed.
Orthographic reference
For precise work, front/side/top orthographic views set up in the viewport let you model against them directly. Photographs have perspective and are misleading used this way — a photo used as an orthographic backdrop bakes its lens distortion into the model.
Reference for generation
The same discipline applies, differently. When generating:
- A good reference photograph produces a good image-to-3D result, and the criteria are the same as for modelling reference — clear, evenly lit, undistorted.
- A prompt is a description of reference you do not have. The more specifically you can describe the object, the closer the intermediate image lands.
- Several consistent views feed multi-view reconstruction directly.
The block-out stage still applies too: check the generated result at scale, in context, before investing in cleanup.