AI image generation
As I’m creating the https://greenseedling.eu site I need a lot of images to make the site more pleasing. These images are used to create a nice mood and feel for the site and don’t need to be of real things. They provide an atmosphere and make boring poges full of text more pleasing to the eye.
One could go to some stock photo site and scroll through thousands of images to find the one that suites your needs. And once you found that image you have to pay the creator of the image to get the right to use the image on your site. The problem is that this cost can add up while these images are just a fill-in.
Of course you could say that I just need to whip out my camera (or smartphone) and just take pictures myself. Of course that would be financially more interesting but getting the right image in the right setting is pretty hard to do. It takes a lot of time to get the right image.
So bring in the modern photography altenative : AI.
A lot has been written about the fact that these AI generated images are based on a training that is done using images taken by real artists. Without these artists there would be no taining for the AI and no generated images. All that is true, but the jury is still out on how this dilemma can be solved. Fact is that photography is changing a lot due to the use of smart phones and computer generated images.
So what proces do I use to generate images?
For the moment I avoid using paying services. At this point in time the generated images are not that good. You can still see they are AI generated and some things are really hard for AI to get right, like fingers and technical things. Try to get AI to generate an perfect image of a sailboat or a plane. Chances are you’ll get something completely bonkers. Once these kinks are ironed out I’ll gladly pay for AI generated images. Some part of the proceeds should go the people that helped train the AI, but as the AI can not tell what images where used when generating a new image, this is not an option.
I looked around adn tried out few AI images generators and got pretty good results using ChatGPT, Copilot and Flux Pro. All of these have a free version with a limited number of images that can be generated in a certain amount of time. If you get your prompt (the command given to the AI to generate the image) right, you generally get the image you are after within the limit that is set. Flux Pro alows you to generate 20 images per 6 hours. Good enough for me.
To get a good prompt you have to describe the image you want as good as possible. To get some isprirtion you can look at prompts that were used by people that generated a similar image. Flux Pro alows you to see the prompts used.
But prompts are pieces of text, adn what software is good are processng text? Large Language Models of course. This is where I use ChatGPT to create a prompt. So I give it something like this.
Generate an photo realistic image like it’s taken with a modern high end digital camera. The image is of a victorian living room with a couch, a coffe table and a dresser. Soft light is coming though the cutains in fron of an open window during golden hour. On the dresser and the coffee table are plants in small pots. In the corner at the right side is a large plant in a large pot. From the ceiling hangs in vistorian chandelier with its lights on. A woman is sitting on the couch reading a book. She is wearing a long green dress.
So I describe the scene as good as I can, mentioning all the details I want. At this point there is no real need for adjectives as we let step 2 do this.
First we let ChatGPT know that we do not want it to generate the image, but we want it to improve on the prompt using the following command:
improve the following prompt for generating an image in an image genarator. Do not create the image yourself
Then we copy the prompt we created above and let ChatGPT do its thing. We get the following result:
Create a highly detailed, photo-realistic image as if taken with a modern high-end digital camera. The scene is a cozy Victorian living room with rich, ornate furniture. A vintage couch is positioned near the center, where a woman in a long, flowing green dress sits, reading a book. The room is bathed in soft, golden-hour light streaming through lace curtains from an open window, casting warm shadows across the space. A classic wooden coffee table sits in front of the couch, adorned with small potted plants, while a matching antique dresser is placed nearby, also holding small plants in elegant pots. In the right corner, a large potted plant adds a lush, natural element to the room. Overhead, a beautiful Victorian chandelier hangs, its warm lights turned on, casting a soft glow that enhances the room’s nostalgic charm.
As you can see, ChatGPT added quite a few adjectives and added a few more descriptive sentences. Check the content and change where necessary. Of course you can ask ChatGPT to rewrite the prompt if needed. Now we copy this prompt and paste it in Flux Pro. I then set the number of images to generate to 1 or 2 just to get the feel of what Flux Pro does ith the prompt. I also set the image aspect ratio to 16:9. If needed I use other aspect ratios of course, depending on the application of the image. I don’t change other settings. I then create the image and get the following
This image is already quite good but I need the plants to be more prominent. The woman is the focal point and not the plants. The image is to be used on a site about plants after all. So I change the prompt to this:
Create a highly detailed, photo-realistic image as if taken with a modern high-end digital camera. The scene is a cozy Victorian living room with rich, ornate furniture. A vintage couch is positioned at the left, where a woman in a long, flowing green dress sits, reading a book. The room is bathed in soft, golden-hour light streaming through lace curtains from an open window, casting warm shadows across the space. A classic wooden coffee table sits in front of the couch, adorned with a mediam sized potted plant, while a matching antique dresser is placed nearby, holding small plants in elegant pots. One of the plants has long branches flowing down the dresser. In the right corner, a large prominent potted plant adds a lush, natural element to the room. Overhead, a beautiful Victorian chandelier hangs, its warm lights turned on, casting a soft glow that enhances the room’s nostalgic charm.
I moved the couch to the left and changed the description of the plants. The result is this:
There are some things that got ignored, like the request for an open window, but overall this is a very pleasing image and serves the purpose.
What do you think?
