Yes. AI can turn a product idea or a rough sketch into concept images. Text-to-image diffusion models, and workflows that turn a sketch into a text prompt and then into images, produce a spread of photorealistic product visuals in far less time than manual rendering takes. Those images are an inspiration aid, though. They give a visual representation of a product idea. They are not functional prototypes, and the models behind them do not understand physical constraints. This is also what separates them from AI generative design in the engineering sense, where an algorithm searches for geometry inside the constraints and simulations an engineer defines.
Is there any AI for designing product concepts and prototype visuals?
Diffusion models are the main generative architecture for visual content. They let people without pixel-level crafting skills generate and edit images from a text prompt or light guidance. In product work, text-to-image models such as DALL-E and Stable Diffusion give designers a visual version of an idea right away and help them explore diverse directions. The MIT DeCoDE Lab’s Sketch2Prototype paper presents this as a faster way to do looks-like prototyping, because designers can generate and iterate on options quickly instead of rendering each one by hand.
A 2023 GUI-prototyping study (UI-Diffuser, built on Stable Diffusion) reports a UI image in seconds that roughly meets the brief. The same authors found missing details and ignored components on closer inspection, and concluded the outputs suit inspiring designers better than serving as working prototypes. That study is about app screens, not physical products.
In engineering design, these tools serve mainly as inspiration during concept generation. A CHI 2024 study examines how AI image generators affect design fixation and divergent thinking. It notes that researchers speculate that generative AI will help inspire designers because it can produce abstract and varied stimuli. We read that as the right job description for concept images: they widen the set of directions a team considers before anyone commits to geometry.
From a sketch or a sentence to a concept image
Under the hood, a text-to-image model encodes the prompt into numerical token embeddings, in UI-Diffuser’s case with CLIP. It then starts from a noisy image and runs a step-by-step diffusion process in which each step adds information that matches the text.
A sketch can enter the pipeline in two quite different ways.
The first is direct conditioning. ControlNet adds auxiliary network modules to a pre-trained diffusion model, so the model can take sketches or spatial layouts as extra control inputs without retraining the base model. The UI-Diffuser work used it this way: ControlNet clones the denoising network into a trainable copy that learns the layout condition, while the original keeps what it learned in pre-training.
The second is sketch-to-text expansion. Sketch2Prototype uses GPT-4V to describe a hand-drawn sketch in words. DALL-E 3 then generates a set of more detailed images from that description, and the chosen images go through image-to-3D generation and postprocessing before being printed with additive manufacturing.
Each route gives you something different. The Sketch2Prototype authors argue that ControlNet and T2IAdapter give users so much control that the images stay too close to the original sketch, so less of the design space gets explored and the designer fixates on the first concept. Going through a text description lets the image model expand the idea, which gave them a much more varied set of prototypes. Our judgement follows from this. Use text expansion when the point is to break out of a first idea. Use sketch or layout conditioning when the form is already agreed and you want variations on materials, colour and styling around it.
In a study of product design-space exploration with Stable Diffusion on Leonardo.AI (“Prompting for Products”), participants used a global mode that regenerates the whole image from a prompt and a local mode that edits a selected region through prompting, masking and erasing. Global-mode prompts averaged 13.4 words and local-mode prompts 3.9 words. In local edits, 59.4% of prompts targeted feasibility, against 19.8% that targeted novelty. In other words, in this bike-design task, local-edit prompts targeted feasibility more often than the other goals.
AI generative design versus generative AI imagery
Autodesk distinguishes the two. In its definition, generative design is a mathematical process that runs many precise real-world simulations to reach optimal solutions for a defined problem. Generative AI produces novel outputs from pretrained neural networks that are “probable, albeit unvalidated”. Generative design works by defining objectives, constraints and a design space, then using algorithms (AI-driven or not) to generate solutions within them.
Neural Concept places generative design early in the process, where it proposes geometry, while topology optimisation refines a geometry that already exists. AI-enhanced generative design explores solutions within the manufacturing method and material layout constraints. Manufacturing constraints in that kind of tool come at a cost too. Autodesk’s Fusion documentation says that applying them limits the freedom of shape generation and typically gives less performant designs.
| Generative AI prototype illustration | Engineering generative design (e.g. Autodesk Fusion) | |
|---|---|---|
| Input | Text prompt, sketch, reference image | Objectives, constraints, design space |
| Output | 2D concept images; optionally a 3D-printable looks-like prototype | A set of optimal design solutions for a defined problem |
| Validation | None built in; outputs are probable, not validated | Simulations run for a defined problem |
| Manufacturing awareness | Not trained on physical constraints | Manufacturing method can be set as a constraint |
| Best stage | Early ideation, stakeholder communication | Early design phase, once objectives and constraints are defined |
The two are complementary, not competing. An illustration can help a team decide which product to pursue, and a generative design study can then find a structure that meets that product’s loads and manufacturing method.
What can a concept image not tell you about dimensions, materials and manufacturability?
A concept image cannot guarantee any dimension, quantity, material property or manufacturing feasibility, because nothing in the generation process enforces them. Each limit shows up differently.
Numbers in a prompt are not constraints
Toyota Research Institute researchers tested whether text guidance can specify quantitative engineering constraints. Asked for “a house with two windows”, a text-guided diffusion model drew a house with windows but ignored the number, and it misread “a car with a drag coefficient of 0.5” completely. The vagueness goes further than numbers. The “Prompting for Products” authors point out that for a prompt like “ergonomic, sleek and modern mug”, it is unclear how the model maps those words onto the image, and they call the tools insufficient for generating reliable product designs. A wall thickness or bore diameter typed into a prompt is not enforced as a constraint.
Proportions and materials drift when there is no geometry
According to Formas (2026), standalone generative platforms work from a flat image or a prompt without reading 3D geometry, so proportions shift and layout logic does not survive. Tools that read actual geometry preserve design intent much better than tools working from a picture. Even so, AI rendering approximates a photorealistic image instead of simulating light across verified geometry. Formas calls it unsuitable wherever material accuracy or regulatory compliance matters.
Turning the image into 3D loses detail and often breaks manufacturability
Sketch2Prototype’s authors call the image-to-3D stage their biggest limitation. Design details get lost there, and the meshes are often non-smooth, fragmented and sometimes non-manufacturable, so they need postprocessing to smooth surfaces, fill holes or remove parts that cannot be made. Uneven surfaces and holes can remain even after postprocessing, which suggests NeRF-based reconstruction suits visualisation better than 3D printing. Control is thin as well. Users can edit the text description or inpaint the images, but at the mesh stage their only control is postprocessing, which is limited to removals and minor surface edits. The Stanford-led diffusion survey adds a warning that matters for reviews: models that output geometry and texture together can hide low-quality geometry behind plausible textures.
Infeasible by default, and convincing anyway
Text-to-image models are trained on images and text without any understanding of physical constraints. As a result they often produce creative but clearly infeasible concepts, which keeps them stuck as inspiration tools in engineering design. The subtler risk is perception. “Prompting for Products” found that feasibility-oriented and aesthetics-oriented prompts were positively correlated. The authors connect this to the aesthetic-usability effect, where people judge attractive designs as more usable even when the function is unchanged. We think this is the failure mode to watch: reading a polished render as a validated design because it looks finished.
Pulling concept images closer to feasibility
On the engineering side, one answer is to condition generation on CAD images that carry real constraints. The “CAD-Prompted Generative Models” study fed CAD images into Stable Diffusion 2.1 alongside the text prompt and got visibly more feasible bike designs. It also found a general trade-off between feasibility and novelty. With the CAD prompt weight kept around 0.35, feasibility improved significantly while novelty stayed on par with text-only prompting. Underneath that sits a further trade-off: the more weight the image prompt gets, the less faithfully the model follows the text, and vice versa. The authors suggest matching the setting to the design stage: text-only prompting for the most novel images in early inspiration, CAD weights of 0.51 to 0.83 when exploring feasible alternatives later.
A second route puts physics-based guidance into the denoising loop. The Toyota Research Institute work interleaves drag-minimisation steps, driven by a surrogate model trained on CFD-labelled vehicle renderings, with the denoising steps of Stable Diffusion. The results stayed coherent as vehicles, with changes ranging from surface smoothing to adding a roof. The authors are clear about the limits, though. A 2D rendering does not determine a ground-truth drag coefficient, so the surrogate’s error cannot be measured. In one case the surrogate predicted very different coefficients for similar outlines, including a physically invalid negative value. They name moving from images to CAD or 3D models as the natural next step.
Building a conditioned pipeline like this is part of Generative & Agentic AI R&D work.
A checklist before an AI concept image leaves ideation
Run these checks before a generated concept becomes the basis for engineering work:
- Numbers: every dimension, count and ratio that matters has been re-specified in CAD, not taken from the image or the prompt.
- Proportions: if the image came from a standalone generator working from a flat picture, the proportions have been checked against a reference model.
- Materials: finishes and materials shown are treated as styling intent, with real specifications chosen separately.
- Mesh: any image-to-3D output has been inspected for holes, fragments and parts that cannot be made, and is labelled looks-like only.
- Texture versus geometry: the 3D form has been reviewed without textures, so plausible surfaces cannot hide poor geometry.
- Feasibility review: someone has judged whether the design can be built, separately from whether it looks good, because people perceive attractive designs as more usable.
The useful question about a concept image is which of its features you are willing to carry into CAD, and who checks each one.
FAQ
Does Autodesk have AI?
Yes. Autodesk offers generative design tools, including the Generative Design workspace in Fusion, and its own definition says generative design may use AI but does not always. Autodesk keeps that algorithmic, simulation-driven approach separate from generative AI, whose outputs it calls probable but unvalidated.
What is a generative process in design?
In engineering generative design, the process means defining objectives, constraints and a design space, then letting algorithms generate many solutions within those limits. In generative AI imagery, it means a model producing new images from patterns learned in training data.