Understanding the difficulty of selling AI art

A short statement: All links I provide have no affiliation with me.

AI art has been going viral for quite some time. The technology driving this trend progresses quickly and firmly. But selling AI-generated art is not easy. In this article, I’d like to share my experience of selling AI art as an amateur, along with some online resources I collected. Hopefully, it can be helpful for you.

Incredible AI artwork made by online communities

Hi! My name is Xiang Zhang. I’m a freelance programmer. Earlier this year, my understanding of AI-generated images was still in GAN’s era. When I saw someone using Stable Diffusion generate stunning beautiful images, I couldn’t believe the advances this field had achieved. I couldn’t help wondering if I could do the same and sell the images in some marketplace.

Soon, I found the famous open-source Stable Diffusion web UI application from Automatic1111 and a couple of excellent Youtube channels teaching people how to use it. Links are below:

• Stable Diffusion web UI
• Sebastian Kamph’s Youtube channel
• Vladimir Chopine’s Youtube channel
• SahiNama’s Youtube channel

Automatic1111’s web UI is so interesting and powerful. It makes intimidating large deep-learning models friendly and fun to use. Stable Diffusion requires GPU to run efficiently. I have a GPU with 6GB VRAM, so I can continue.

In about a month, I learned:

1. How to install Stable Diffusion web UI and extensions locally.
2. How to construct a prompt properly.
3. How to do image-to-image and inpainting.
4. How to use ControlNet to control a character’s pose.
5. Where to find a custom model, Lora, embeddings.
6. A workflow from an idea to a high-resolution image.

The additional resources I learned from are below:

• OpenArt’s prompt book
• Stable Diffusion Art Website
• AI Art Alchemy’s video on prompt construction
• CivitAI, an AI model-sharing website

Then I got down to the business part. After a little research, I found people could sell artwork as digital downloads, or join print-on-demand(POD) platforms. These could bring passive income. Other options include starting a commission gig on freelance platforms.

I started from a POD platform. I experimented with different prompts to generate images, uploaded them to shops, edited products, and promoted them on social media. It looked like a perfect plan, however many hidden problems were gradually exposed.

First, Stable Diffusion 1.5 is trained on 512×512 image datasets. The generated images have to be similar sizes. If you give it a larger size, it’s prone to generate weird things, such as two heads, or repeat the same subject multiple times. So a common workflow is to generate low-resolution images first, then upscale them. I have tried different upscaling methods so far and there is only one that I’m satisfied with. That is Hires-fix. But it costs computation resources and time. My potato PC can’t generate images of more than 2K with this method. Other methods like SD-upscaling or ControlNet-tile-Ultimate-SD-upscaling either produce weird textures or generate inconsistent random levels of details, not to mention tile seams.

Zoom in on details of the original(left), SD upscaling(middle), and Hires-fix(right). SD upscaling may treat subjects of different depths as one piece of texture. As you can see, Her eyelashes are almost embedded into her eyelids(middle), while Hires-fix(right) looks more natural.

Second, a generated image is not perfect. A small error can cause a big headache. I often found myself inpainting images to remove things. Most of the time, I wish I were a real painter so that I could use a brush instead of luck. Because of this, I learned some basics of photo editing. But still not enough. To provide high-quality design, I may also have to know how to bend text, vectorize images, create different effects, compositions, etc. I will have to be a real designer or artist to finish up an artwork.

An example of inpainting frustration. Can we just hide all those weird stuff? Thank you.

Third, I don’t know what people would buy in advance. There is no guarantee that my work could be sold. This process is trial and error. At the time of this writing, I managed to upload 30 designs to my POD shop. Some of them are rabbit and cat painting. Some of them are flower patterns. I thought about T-shirt design, but it’s beyond my skill set, all I can produce are small rectangular images.

At last, my sales in three weeks are zero.

A collection of my POD shop’s artwork generated by Stable Diffusion

In conclusion, although the learning process of AI-generated art is full of fun and joy, monetization still faces difficulties:

1. Images are small in size, upscaling can be a problem.
2. They are not perfect and still need enhancement and editing.
3. They need to meet people’s preferences to be sold.
4. This additional one is a heads-up on legal issues. Some licenses should be checked, including not only the content(IP infringement) but also the models.

An example of usage restriction of an AI model

That’s my story. What do you think of selling AI art? Or any suggestions? Please leave comments below.