Image to 3D model

A photograph contains no scale. That single fact explains almost everything about what image-to-3D can and cannot do, and it is the thing most tools in this space are quietest about.

Point a camera at a bracket and the image records shape, proportion and arrangement. It does not record that the part is 60 mm long, because the same pixels would result from a part 6 mm long photographed closer, or a part 600 mm long photographed further away. No amount of model capability recovers a number that was never captured.

This is not a limitation to work around. It is a division of labour. The image supplies the form; you supply the numbers; together they produce a model that is both the right shape and the right size.

What a reference image actually contributes

Attach an image in TextoCAD and it goes to the model alongside your prompt, resolving the ambiguity that words carry badly. Some things are genuinely tedious to describe and instant to show:

• Overall formwhen the part is not a name you can reach for. "A sort of tapered bracket with a stepped foot" is three attempts at a sentence and one glance at a picture.
• Arrangement. Which side the flange is on, how the cutouts sit relative to the mounting holes, which way the hook faces.
• Proportion. Not absolute size, but the relationship between parts: that the base is roughly twice the height, that the slot runs most of the length.
• Style on decorative work, where a described shape and an intended shape can diverge a long way.

What it does not contribute is any dimension whatsoever. That stays your job, and it is why the prompt still matters.

Taking a usable reference image

The difference between a good and a poor reference is mostly about removing things the model has to interpret.

• One part, filling the frame. A workbench with six objects on it forces a guess about which one you mean.
• Plain background. Sheet of paper, plain worktop, anything without pattern. Contrast between part and background does most of the work.
• Square on, not three-quarter. A straight-on view of the face that matters beats a more attractive angled shot. If two faces matter, take two photographs.
• Flat, even light. Hard shadows read as features. Diffuse daylight is ideal, and a direct flash is the worst option.
• Put a ruler in the frame. It will not be measured for you, but it keeps you honest when you check the result against the original.

Sketches often work better than photographs

A hand sketch has an advantage a photograph cannot match: it contains only what matters. No background, no lighting, no reflections, no perspective distortion. Every line you drew is a line you meant.

The strongest version is an engineering-style sketch with the dimensions written on it. Draw the part roughly to proportion, mark the overall sizes, hole diameters and thicknesses directly on the drawing, photograph it flat, and attach it. You have now communicated shape and numbers in a single image, which is exactly the split this whole page is about.

This works well for parts you are copying from something physical. Measure the original with calipers, sketch it, annotate it, generate it.

How this differs from photogrammetry and mesh reconstruction

Two other technologies get grouped under "image to 3D" and they do a different job.

Photogrammetry reconstructs a surface from many overlapping photographs taken around an object. With enough coverage it produces an accurate mesh of what is really there, and it is the right tool for capturing a statue, a rock, a face. It needs dozens of images, it struggles with shiny and featureless surfaces, and it gives you a scan.

Single-image mesh generation takes one photo and infers a plausible three-dimensional shape, guessing the parts it cannot see. Impressive for concept work and game assets. Not a measurement.

Both produce meshes: triangles with no editable dimensions and no STEP export. If your goal is a decorative object or a digital asset, they are the right tools. If the part has to bolt to something, the thing you need is a solid, and AI CAD generators compared covers that distinction in full.

The practical workflow

Image plus stated dimensions, then correct on the sliders. Concretely:

Why results disappoint, and how to fix it

Nearly every poor outcome traces to the same cause: the image was treated as a replacement for the prompt rather than an addition to it. An image with a one-word instruction gives the engine a shape and no requirements, so it invents the requirements. They will not be yours.

The rest, in rough order of frequency:

• No dimensions given. The most common by a wide margin. Say the size.
• Busy background. Reshoot against plain paper. It takes twenty seconds and changes the result.
• Angled photograph. Perspective makes a rectangle a trapezium. Shoot square on.
• Hidden features. A hole on the far side is not in the image. Describe it or photograph that side too.

When the shape is right and the numbers are not, do not re-prompt. Move the sliders. Correcting a thickness on a slider is faster than describing it again, and it does not spend a generation.

Where to go next

The prompt half of this is covered properly in CAD prompt examples, and how to convert text to CAD walks the whole workflow. If the part is destined for a printer, design for 3D printing has the clearances to state in the prompt. The gallery shows the prompt behind every model, which is the fastest way to calibrate how much detail is enough.