The State of AI Fashion Photography 2024

We started working on AI fashion photography about a year ago. But the technology was just not good enough then. About one out of 100 generated images were usable—and that’s if you don’t look too closely.

Over the past year, AI has gotten much better and people have come up with various approaches to generate better AI fashion photos (including ourselves!)

We believe AI fashion photography is now ready for businesses to use.

In this blog post, I want to share various approaches to AI fashion photography, how businesses are choosing between them, and what we believe is the best approach for businesses.

How to assess AI fashion photography tools

Before we dive into the four approaches to AI fashion photography, let's first answer the question:

How do you know which approach is the best for you?

After speaking to hundreds of our Pebblely customers, we believe there are three criteria for assessing AI product photography tools for businesses:

1. Accuracy: The product, in this case, clothes, in the generated images must match the actual design because you are selling the clothes. You cannot show something other than what you are selling. While it seems silly to say this, many AI tools generate similar but not the same clothes.
2. Cost: If you want to use AI to replace traditional photo shoots, the AI should cost drastically cheaper—or at least 10 times cheaper. If it costs only slightly cheaper, it is rarely worth the hassle of adopting a new tool and changing your processes. You should also consider the “labor” costs. If you want to use AI, it should not require significant manual work, such as learning technical details and Photoshopping.
3. Speed: Another value that AI promises is fast turnaround. Instead of spending weeks to months on a traditional photo shoot, you can generate images in seconds. Even if you don’t get what you want on the first try, you should be able to easily generate more images immediately.

It is also important to consider how photorealistic the generated images look. But most AI images are realistic enough and will likely become even more realistic in the coming year. So I take photorealism as a given.

Let's keep these criteria in mind as we explore the state of AI fashion photography.

The state of AI fashion photography

To help you navigate the space, here are four common approaches to AI fashion photography nowadays:

1. Combine 3D designs with AI models
2. Generate images with reference image(s)
3. Swap models for existing images
4. Turn product photos into on-model photos

1. Combine 3D designs with AI models

If you use 3D fashion design software to create your designs, you can now put the 3D renderings onto AI-generated fashion models.

The advantage of this approach is accuracy. Because you know the exact dimensions from the 3D designs, you can fit the clothes on different models accurately. If you care most about the details, this is likely the best approach for you.

But 3D fashion design software is expensive. Tools like VStitcher and CLO don’t even show their price on their website. A fashion CEO told me it can cost up to hundreds of thousands of dollars every year. And you have to learn how to use the 3D software first. Without 3D designs, you cannot use this approach.

Thanks to the recent developments in AI image generation, we now have several cheaper, easier-to-use, and faster options.

2. Generate images with reference image(s)

General AI image generators do not allow you to put your clothes into images directly. But you can use an image of your clothes as a reference.

• In Midjourney, you can use images as part of a prompt (also known as image prompts). See the example above.
• With Stable Diffusion, you can input an image (also known as image-to-image or img2img for short) or use several images to get more custom results (also known as DreamBooth or, some would call it, training a custom AI model).

The advantage of this approach is it is much more affordable. Midjourney’s cheapest plan is just $10/month. If you are willing to roll up your sleeves and set up Stable Diffusion on your computer yourself, you can even generate images for free. In the open-source world, the most popular options are A1111 and ComfyUI.

The dealbreaker is that your clothes will be distorted. As you can see from the Midjourney example above, the generated images do not have the same print as the input image. The texture and fit will also be incorrect. To deal with this, you could Photoshop the original design onto the images or retouch the images to change the texture.

Also, while tools like Midjourney are marketed to be easy to use, it takes much practice to get good results. Using Stable Diffusion requires you to be at least a little technical.

3. Swap models for existing images

If you want your clothes to look the same as your photos, model swap is an option. It takes an existing model photo (or even a mannequin photo) and replaces the model with an AI fashion model. (Technically, this uses Stable Diffusion’s inpainting and ControlNet.)

The advantage of this approach is your clothes will look the same as your original photos, which is essential for commerce. This approach is usually used by brands who want localized models for the different regions they are selling to but do not want to shoot with multiple models. Or they simply do not like the model in their photos. With AI, they can take an on-model photo and generate more photos with different models and even different backgrounds.

There are two main limitations to this approach. One, you first need to have on-model photos. This doesn’t work if you only have photos of your clothes. Two, all your on-model photos will have the exact same pose.

4. Turn product photos into on-model photos

Finally, you can put photos of your clothes on AI fashion models. With just a simple flatlay image, you can generate AI fashion models of different ethnicities wearing your clothes in multiple poses and backgrounds. Most importantly, the text and patterns on the clothes will be preserved in the generated images.

But this approach doesn’t consider the actual fit or size of the clothes. For example, a loose-fit shirt will always fit nicely on the models. (But, arguably, accurate model photos still have not solved sizing and fit problems for online shopping.)

Also, the AI might extend parts of the clothes unnecessarily, such as adding sleeves to a sleeveless top. As this issue doesn’t happen all the time, generating more images will give you usable images. We also foresee this issue going away within the next year (and are actively working to make it happen!)

How brands are choosing between these approaches

From our research and conversations, here are some patterns we noticed:

Big brands that are very mindful of scrutiny from the media and public prefer to start with the 3D design approach, especially if they are already paying for the expensive 3D design software. This approach gives them the level of accuracy they require while still providing some cost-savings, compared to traditional photo shoots. An example is Levi’s.

More forward-thinking big brands, such as Casablanca and Louis Vuitton, are experimenting with the other approaches. They usually start with a campaign to show that they are keeping up with the AI trend. Some would Photoshop generated images until they are suitable—just like how they would already Photoshop photo shoot photos.

Small-to-medium brands are generally more open to using AI-generated on-model photos—even if the results are not perfect yet—because of the enormous cost-savings and the potential to increase conversion rates. Luxury knitwear brand Sheep Inc. actually launched a campaign with only AI-generated model photos!

Our approach at Pebblely

We want to simplify awe-inspiring but intimidating AI technologies so that businesses around the world can use AI to grow.

We built Pebblely Fashion with the fourth approach in mind because here are our beliefs:

1. Turning product photos into on-model photos provides the most value to businesses because you can skip model photo shoots entirely. Also, the downsides to this approach is more trivial (e.g. wrong fit, instead of wrong prints) and some issues can be resolved by generating more images. We are also working to improve our AI to minimize these issues.
2. For commerce, the prints, logos, and text on the clothes in the generated photos must be the same as the actual design. We trained a specialized AI to do this.
3. The cost should be significantly lower than traditional photo shoots. This is not just to help big businesses save costs but to also allow small businesses without much budget to finally have on-model photos. The cost per image for our Pebblely Fashion plan is about $0.10 or lower.
4. The tool should be easy to use and does not require technical setup or training. Just by describing the model and background you want, you can generate a variety of on-model photos. See our demo here.
5. You should get results fast. Each image takes only about 20 seconds to generate. Also importantly, Pebblely Fashion is entirely self-served. You do not have to jump through sales demos and approvals to even try it. You can easily try different things and iterate with the Pebblely Fashion app yourself. (Of course, we are happy to help if you have any questions. Just email us!)

If this matches what you are looking for, we would love you to try Pebblely Fashion and let us know your feedback.

P.S. We are also looking to partner with fashion brands—big or small—to improve our AI and help you generate better images. If you are interested, feel free to reach out.