AI assisted design

We are currently evaluating AI technologies as a tool for streamlining our development processes. One of the possible AI use cases that we see is helping to rapidly produce ideas in the early stages of a project as part of an AI assisted brainstorming session.

This is the results of one such session which took place to help create ideas for an enclosure for a hand-held console. This enclosure required certain features and the AI was being used to help generate enclosure variations using these features.

The AI system used generates four images in response to a user supplied prompt describing the requirements.

Generating images with the prompt "handheld game console blank screen".

This generation immediately shows the common problem when using AI for image generation of regurgitating the training data and displaying images of the images originally used to train the AI model. In this case it's obvious that images of a Nintendo Game Boy were used in the training data and have been reproduced almost verbatim.

Here's a Game Boy image for comparison. In this case these images are easily recognisable which would hopefully make this issue easily detected but the use of designs generated by AI in a product could lead to copyright issues.

For the next attempt try generating images with the prompt "rounded plastic rectangular box with 10 differently coloured buttons and blank LCD screen" to avoid game consoles being reproduced.

These all look very much like childrens toys, which could be the desired result, but one of the generated images has no LCD display which was a requirement in the prompt text.

To make the images look less childlike try generating images with the prompt "rounded plastic rectangular gothic console with 6 blue 1 green 1 yellow and 2 red buttons and blank LCD screen".

These look interesting although the bottom right image does look like a tombstone.

Generating images with the prompt "rounded plastic rectangular gothic console with 10 differently coloured buttons and blank LCD screen".

Only two of these images has an LCD screen contrary to the requirements.

Generating images with the prompt "rounded plastic rectangular gothic look box with 10 differently coloured buttons and blank lcd screen".

It's hard what to make of these images as none of them are useful in the slightest; indeed one of the images looks like a plastic box loosely filled with clothes buttons.

Other sessions that are not shown here have demonstrated that currently some common shapes are not understood by the AI. For example, a prompt to generate an 'L' shaped object merely resulted in rectangular shape being generated.

It's very hard to predict the images generated from the AI in response to prompts. The small representative sample shown here varied from copyright infringing images to random images missing essential features that were specified in the generation prompt. Currently we are looking at the possibility of using these images for AI assisted brainstorming sessions only until this technology improves and these sort of problems are resolved.