Can AI help create technical illustrations?
AI is a big topic, and we have previously looked at how it impacts product information, but here we want to look at it from the perspective of technical illustrations.
Lots of images we see daily are now AI generated, be it the cover photo for the office Christmas party or product placement images from brands we follow on social media. But what are the use cases outside “fun”? Can AI help create technical illustrations?
Let's start by looking at the current potential:
We know AI can generate images with lots of content and can easily adapt it with prompts like “make the background a cityscape” or “add more people”. But in the case of product information and technical illustrations, it’s often the details that matter, and AI is not good at the details.
AI also is just creating images from reference datasets, so they often make images with six fingers on someone’s hand, or a cable for a wireless product. Text in images is also not something AI does terribly well at the moment.
We thought we would conduct a simple test:
We pitted one of our illustrators from our sister company Sigma Technology Hungary against an AI tool. We thought we would keep it simple, so we picked a Microsoft keyboard as the product to illustrate and Microsoft Copilot as the tool.
But we didn't pick a standard keyboard, as we wanted to challenge our illustrator as well. (Straight lines are far too easy, surely?) So we picked the Microsoft Ergonomic Keyboard.
As is common for illustration tasks, a reference image was difficult to source with a high resolution. Obviously a CAD file would have been better input, but sometimes in tests, as in work, you have to work with what you've got.
The illustrator:
The illustrator didn’t need much direction, just told what sort of style (rendered, line, and so on) we wanted and from what angle. We then even got some questions back: format, size/resolution, file type, background or transparent. We asked for a simple line drawing.
The AI tool:
As you can see, an illustrator is good at details. We don’t need to go into a discussion about which LLM or AI tool is best for images, but we thought a Microsoft product from a Microsoft tool should be possible.
Clearly, it wasn't. The images created by Copilot were of random generic keyboards, even fine-tuning the prompting didn't give better results. It was hit-and-miss if a new prompt improved on the last image or not.
So, we had to use another tool, one where we could add more reference images. However, even after several prompts to fine-tune the output, we still didn't have a usable file.
Some of the output was close to what we would want. But it isn't actually a line drawing. It's a pixelized or stylized image. So it isn't something we would be able to reuse later and is not something that could be zoomed in on either.
OK, so it seems that AI can’t help with creating technical illustrations at the moment, but what can it be helpful for within technical illustrations?
AI is very good at generating ideas, people are very good at decision making. So asking for sketches and drafts is a good way to save you time and have more ideas to work from.
AI can help with setting moods and changing elements within images quickly. So perhaps if we create a technical illustration of the keyboard and then ask AI to “put it on a desk”, we will get a final image, or several final images quicker.
The future:
AI lost this battle, but it isn’t totally useless in this area. Together with Sigma Technology Hungary , we are looking at more specific usage and training of AI to see what is possible. Watch this space.
Information & Communication Specialist at Sigma Technology Information
3 个月"No" is clearly the answer at the moment, but we have some idea of potential applications (adding/changing backgrounds etc. as written in the article). I'm looking forward to seeing what more Nándor and crew can get up to in their further investigation ??
Learning Lead at Sigma Technology Information
3 个月Yeah, I agree. So far I could only make use of “fun” or very “general” AI-generated images. It’s really bad with details and specific instructions. Also, most AI images have a weird style that is really noticeable.