Hacker News Reader

Comment by tux1968 | Hacker News Reader

tux1968

You know a lot more about this than me, but one potential counterpoint though is Ikea already using ~75% computer generated images in their catalogues. When the technology is mature enough, it provides a huge amount of flexibility compared to a photo shoot.

View on HN

From my own attempts at hobbyist 3d modelling, computer generating a human with clothes is far harder than furniture. Looking through the computer generated images for IKEA almost all of them are hard body (solid objects) with very few soft objects (clothing, tablecloths, etc) and nothing that is alive. Creating realistic clothing renders is quite difficult, and it gets even more difficult when a lifelike human needs to be rendered as well.

Consider the incentives fast fashion vendors have, and it seems clear that completely AI generated product photography is inevitable.

It's expensive and slow to hire a photographer and a studio and a model, and you have to ship the clothes there ahead of the launch on your website to have them spend all day getting in and out of different outfits while stylists keep their hair tidy, then the photographs go to the art direction team to be photoshopped to match the site aesthetic...

If you can just get someone in the factory in China to snap a photograph of the latest batch of dresses and tops and skirts as they come off the sewing table, then you can just send them into a GAN, and have style-matched 'photographs' generated showing the clothes on a selection of different models, each of whose appearance is perfectly tailored to appeal to different market segments.

You can have high quality creative on your website and in the product feed to Google the same day, and start taking orders before the inventory starts piling up.

Then next week, you can do it again with the next set of designs.

> If you can just get someone in the factory in China to snap a photograph of the latest batch of dresses and tops and skirts as they come off the sewing table,

Just a simple photo of the garment will not tell you how it behaves on the body (how "malleable" it is, how it bends, wrinkles etc.) and how it interacts with the light. Just ask any artist about the nuances of painting clothing materials - it's a big subject in its own right. I suspect that only shooting photos on a in-factory models, in various lighting conditions MIGHT be enough to train the AI.

There’s plenty of training material available - the internet is full of pictures of people wearing clothes.

Yep, but how is that informative for a particular garment that you want to visualize? You can't infer, from a corpus of images of random clothes being worn, how that particular shape and fabric behaves.

What I meant to say was not that the technology was ready today, rather that when it becomes ready it's likely to be used. At least it was in the case of IKEA when the technology was able to meet their needs.

You don't need to have full computer generation to dis-employ a lot of models. Just making modeling a lot more productive would do that. For instance, you could have a "deepfake" approach that replaces the clothing a model is wearing. That would enable a small number of models and a pile of software to generate a hell of a lot of imagery.

The way modeling can work is that you take a photograph of a mannequin with the clothes upon it then digitally add in facial features or skin tone.

Even just being able to shift skin tone, and not show a face will work and will add to profitablily because people do want to know that clothing will work with their complexion

Ehh, this could backfire if it ends up in the uncanny valley realm. Consumers prefer images of models wearing clothing, because it shows the fitment. They may flock to brands that explicitly use real models.

IMO, modeling is one of those things that will never go away because of how cheap this labor is compared to hiring someone to make a software solution. Just like how it is cheaper to stock a fast food joint with a few minimum wage people worked to the bone than a sleek robot with a pricey service contract, even if the latter is sexier.

From what I know about the industry, there's also different categories. I don't doubt that Target or Walmart could use digital models for generic clothing. I doubt that would work for a designer launching a new high end collection. Then there's the experiential side where designers like Oscar de La Renta will have events with live models for women to view their red carpet/wedding dresses. Potentially, software could help designers quickly test and iterate on new clothing ideas, but launch brands likely will launch campaigns around real people.

They were at 75% in 2014. They might be closer to 100% now. Ever since I learned about IKEA's CGI work, I always take more time to browse their catalogs.

https://web.archive.org/web/20141230115206/http://www.cgsoci...

which does probably not really relate to AI. all the chinese amazon-sellers do it as well.

Reply on news.ycombinator.com