Generative AI for Product Design: From Visual Models to Usable Structures

From “Looks Like” to “Actually Works”

Generative AI is increasingly capable of producing images, 3D models, and structural drafts from text prompts. Describe a chair, and it can produce a seat, back, and legs. Describe a device, and it can generate an exterior concept with an enclosure, openings, and controls. This is already valuable for concept presentations, visual creation, and early ideation.

But when the task shifts from “draw an object” to “design an object that can be manufactured, assembled, and operated,” the evaluation criteria must change. Similarity in appearance, rich detail, and realistic rendering do not prove that the design holds up in the physical world. A wheelchair can have four wheels in an image but no structure that lets the wheels turn relative to the frame. A toaster can have slots and a knob but lack the internal mechanism needed for heating, timing, or resetting.

The core problem for generative design is therefore not whether it can generate structure, but whether it can generate structures that are functionally complete, correctly related, and verifiable.

Visual Quality and Engineering Validity Are Different Measures

Traditional 3D generation evaluations often focus on geometric appearance: whether the object’s outline matches a reference image, whether surfaces are smooth, and whether details are rich. These measures suit appearance modeling and visual content production, but they cannot directly answer engineering questions.

A visually realistic model may still have the following defects:

• Missing critical parts: The object has a visible shell but lacks internal components required for function.
• Incorrect connections: Parts exist but are not properly connected, or assembly positions are unreasonable.
• Invalid motion relationships: Parts that need to rotate, slide, or extend are fixed, or they collide when moving.
• Non-manufacturable geometry: Part thickness, clearances, hole positions, or shapes do not meet common machining and assembly requirements.
• Insufficient strength and stability: The structure can be placed in a model but cannot withstand loads or impacts in real use.

“Looks right” covers only part of product design. For real objects, shape, function, structure, material, and manufacturing process are interrelated, and judgment cannot rely on appearance alone.

Function Is Often Delivered by Unobtrusive Parts

For a wheelchair, users notice the frame, seat, backrest, and wheels. But whether it can move also depends on how the wheels are mounted, how the axle connects to the frame, whether the caster wheels can change direction, and whether the brake mechanism can limit motion. Generating an appearance model with four wheels does not prove that these parts have a working mechanical relationship.

A toaster has a similar issue. A bread slot and a control knob can be clearly presented on the exterior, but heating elements, control mechanisms, timing or temperature-control parts, and the pop-up mechanism together determine whether the device can complete the toasting process. With only an enclosure and a user interface, the model may look like a product without having product function.

The same is true for toilets, door locks, folding furniture, and other objects. Hinges, linkages, valves, and locking parts are usually less visible than the exterior shell, yet they perform key roles in connection, force transmission, sealing, or control. Trained only on large numbers of appearance images, a model may more easily learn “what parts usually look like” but struggle to reliably infer “why the parts exist and how they should fit together.”

An Evaluation Framework for Real Manufacturing

To assess whether a generated result has engineering value, we should not assign a single “looks similar” score. Instead, the task should be broken into multiple checkable layers.

1. Component Completeness

First, define what functions the target object must perform, then list the parts required to implement them. The check should focus not only on which objects appear in the exterior, but also on whether necessary support, connection, moving, and control parts exist.

This step can start with a function list. For a movable seating structure, for example, first define functions such as load bearing, movement, steering, and braking, and then check whether the design has corresponding parts. A checklist cannot replace professional design, but it can help reveal the problem of “complete appearance, missing function.”

2. Connections and Assembly Relationships

Parts existing does not mean they can form a whole. It is necessary to check whether connection locations, assembly directions, fastening methods, and spatial layouts are reasonable. Two parts may each have the correct shape yet fail to assemble because hole positions do not match, interface dimensions conflict, or installation space is insufficient.

A generated result should therefore express relationships among parts, not merely provide a set of unrelated geometric shapes. For complex objects, assembly order should also be made explicit: which part is installed first, which connections must be completed before an enclosure is closed, and which parts need space for maintenance or replacement.

3. Motion and Constraints

For objects with rotation, sliding, folding, or extension functions, degrees of freedom and range of motion need verification. Does a part move around the correct axis? Does its movement area conflict with other structures? Does it jam at the end of travel? Are limits or locking mechanisms needed?

Motion constraints are not just animation effects. A model can play a rotation animation without real parts having the shaft, bearing, clearance, and support structure needed to achieve that motion. Virtual motion checks can help expose interference problems, but they still need to be verified against real dimensions, tolerances, and material properties.

4. Loads, Materials, and Stability

Structural design must consider forces during use. Load-bearing parts require evaluation of how loads are transmitted to supporting structures, whether weak connections may break, and whether the center of gravity may cause tipping. The stiffness, toughness, weight, and wear resistance of different materials also affect the design result.

If a generated model lacks sufficient material and operating-condition information, it should not be treated as a validated engineering solution. A reasonable approach is to position the model as a candidate design, then validate key assumptions through computational analysis, prototype testing, or professional engineering review.

5. Manufacturing and Maintenance

“Can be drawn” does not mean “can be built.” Manufacturing evaluation needs to consider processing methods, part count, tool access for assembly, dimensional tolerances, and material availability. For products intended for long-term use, the inspection, replacement, and maintenance paths for wear parts must also be considered.

If a part is completely enclosed by other structures and has no reasonable assembly or maintenance access, the model may be difficult to manufacture and repair even if its appearance is complete. Bringing manufacturing constraints into the design process early is usually more effective than fixing problems after the concept is finished.

Why Models Tend to Omit Functional Structures

Generative models typically learn shapes and patterns from existing data. Visible appearances often appear in large numbers of images, while internal mechanisms, assembly processes, load paths, and failure cases are not easily represented through ordinary images. This uneven data distribution can make models more familiar with the outside of objects without necessarily grasping the causal relationships of internal structures.

In addition, text requirements themselves may be ambiguous. “Make a wheelchair” does not specify the operating environment, load requirements, size range, steering method, braking requirements, or manufacturing conditions. Without necessary constraints, a model can only fill in details based on common appearances. Even if the completed result matches a typical visual impression, it may not meet specific use requirements.

There is also an important distinction: generation proposes candidate solutions, while validation checks whether a solution satisfies constraints. The ability to generate a seemingly reasonable answer does not mean the model has completed physical simulation, structural analysis, or safety certification. Treating generated results directly as production-ready designs confuses these two different stages.

A More Reliable Workflow

In real projects, generative design can be placed within a process of gradual convergence rather than expecting the model to deliver a final product in one step.

1. Define requirements and boundary conditions: Specify functions, use scenarios, size ranges, loads, material constraints, and manufacturing methods. The clearer the requirements, the more basis there is for later evaluation.
2. Generate multiple candidates: Treat model outputs as ideation material and compare different structural approaches rather than viewing a single result as the standard answer.
3. Build a function-to-part mapping: Confirm item by item which parts implement each function, and check the connections and force-transmission relationships among parts.
4. Perform geometry and motion checks: Check dimensions, clearances, collisions, motion ranges, and assembly space, and eliminate clearly infeasible options early.
5. Carry out engineering validation: According to risk and use scenario, use computational analysis, prototype fabrication, load testing, or professional review.
6. Record assumptions and uncertainties: Distinguish verified facts, unverified inferences, and information that still needs to be added, avoiding the mistake of calling a schematic model a manufacturable design.

For products with high safety requirements, automatically generated options especially require professional participation. Human review is not merely polishing a model; it supplements domain knowledge the model may lack and assigns clear validation responsibility for key risks.

Conclusion

Generative AI is lowering the barrier to 3D ideation and concept expression, but between a visualization model and a working product lie functional decomposition, structural design, assembly validation, material selection, and manufacturing testing. Whether a wheelchair can turn does not depend on whether the image contains four wheels, but on whether the parts form a system capable of motion.

When evaluating this technology, we should look at whether it can generate appearance and ask whether the design is complete, the relationships are correct, and the functions are verifiable. Only by distinguishing “looks like” from “can work” can we accurately judge the value of generative design and integrate it safely and effectively into real engineering workflows.