[The NO Prompt Method] MULTIPLE Consistent Characters with Custom GPT & DALL-E

My goal today is to build a story illustrator bot in ChatGPT that will create multiple consistent characters for your story. So once you are set on your story, character details, you can start asking an illustrator bot to put them in the environment and the context without repeating tedious prompt each time. you can discuss with it too how to better structure the composition and finetune your images with natural languages.

I will also share with you my ultimate hack of achieving character consistency. Finally, you can have an illustrator that actually understand the story. So it will create images that present the best details possible to match your narratives.

So let's get started. the process of image generation of this GPT is that when the user send in the requests, the GPT bot will then take into the consideration of the configuration and the instruction and you have at the GPT backend and then generate a prompt of under 400 character limit or about that range to DALLE and then DALL-E will generate an image as the output. And now as I tested, the GPT would not use gen ID and seed number in its image prompt to DALL-E to generate images.

So what you input instruction into the GPT is the only thing that you can control. If you want to know about the secret rules that OpenAI used to prompt ChatGPT with DALL-E. You can find the details in the notes in the description below.

Setting up your character design and our style is the most important step when creating this GPT bot. You can ask ChatGPT to give you suggestions of the character designs, but I have my main character design for my new storybook that I created myself, and I wouldn't want to let ChatGPT to take away all the fun of story design. My first character is a Yoko.

She is eight year old, Japanese girl, dark shoulder length hair with brown eyes. You want to set the age because if you don't, then it would probably give you some full-grown woman when generating that image. And for outfit.

I got a purple hair band with a bow, sleeveless purple dress over white t shirt and brown Mary Jane shoes with white socks. I wanted to look as distinct as possible. And the best way I think to do this is with your outfits.

That's why I specify very specifically how she should look in terms of outfit. I also have Marcus, and to test, I also have a animal. Lucky To maintain the consistency of animals, I would recommend to specify an easy identifiable dog breed.

So a simple word like Corgi has embodied a lot of characteristics DALLE would know. And I also recommend to have dog features with less uneven markings such as spots or colors to decrease The failure rate of different results. when you design these characters, you want to get as specific as possible with the important features.

But at the same time, you want to use as fewer words to describe them as possible. Because when the GPT sends instructions to DALLE it will only include less than 400 character long of prompts. And the more the characters of the prompts, the more they will confuse DALLE and create horror images.

But don't worry, some of the details loss can be easily corrected after the images are created. I will show you later in the steps. also it's important to determine the art style so we can get a more consistent look and feel of the images For me, I am using 3D Pixar animation style as I know this is the style that it has been trained extensively on.

And also I love Pixar. If you want to find your own unique style, you can check out a super helpful resource I found online that can teach you how to get the most out of DALLE and art styles that it was trained on. Click on the notes of the video link in my description below to get the DALL-E Prompt Book Guide.

If you want to dive deeper and you will find a lot of styles It was in this document. After setting up the character design and the art style, let's get into the building of the GPT bot Let's build the bot. So let's go to ChatGPT.

Go to explore under my GPT create a GPT You can build the bot by talking to it and telling it what you want, but it requires a lot of back and forth. and when you place your various request sometimes it would wipe out the previous requests chat with it to let it knows what you like. I find it faster to just go to configure or to input your own information.

So add the name story illustrator and description if you want to. And also input the instruction here. this is the instruction the bot builder created after our chat.

So I prefer to have an instruction written out in a separate place and copy paste. This instruction is from multiple back and forth chatting session with a GPT builder to get the best instruction I think is suitable for my needs. You can take it as a reference or adapt it to your own and I will share this in the notes in the description below.

Okay. So copy and paste. I will quickly walk you through what I got here in my instruction First is the purpose of the bot in a brief mention of the back story.

and second, is this an important instruction of how it should behave. So in here, I said using establish character description and based prompts, the GPT ensures consistency in their appearances, outfits and expressions across illustrations, each illustration request will center on the specific scenes, context, Whenever GPT generates an image, it will always maintain consistent visual style and use the base prompts for subject description for every image generated the GPT will ensure high resolution and high quality Pixar 3D animated film style, and the characters should be depicted with the same visual style proportions and clothing details as the reference images unless specifically requested by a user. The GPT will not alter the set outfits and features of the characters.

Always generate the image in aspect ratio of 16 by 9. This is because I want to make a movie out of the images, So in here I said I want it to follow a formula when you create the image prompt a specific formula that will be used for creating prompts for DALL-E The first element I want is the subject description. I wanted to utilize the base prompts below for the characters All character description in a base image prompts below for each character must be fully included in every image prompt.

and environment description will be included and it will detail the scene setting based on user requests or generate an appropriate background, if not specify. And art style. I want it to consistently using Pixar's 3D animation style and for color and light, I want it to describe the main colors and lighting, I also, as it always include the camera angle and composition to indicate the perspective for the scenes for the reference image usage, as I mentioned, I don't know how much it would affect things, but I still include this as per the GPT builder’s suggestion.

So I just include this just in case. And the base image prompts for each characters, as I have shown you. After adjusting the instruction, come down to knowledge and you can upload files here for any other information you want it to know.

here, I would actually include the images I have how do I come up with these images? after you have the instruction here, you can go back to create and maybe you can even save it and start chatting with it and ask the GPT builder to create some samples for you. for the capability section, you want to click all three of them.

So it would be able to search online and use DALL-E. And for code interpreter, you want to sometimes upload your own reference images directly at the chat bot so it can refer to it, maybe create something similar. and for Action, we don't need to do it for this particular chat bot so we can just leave it.

And I don't really like this image, so I'm going to click it and upload my own image. This published to only me and confirm the first time to test it. I would actually come back to explore and to edit the bot because then if anything goes wrong, we can directly chat with it here to correct.

So let's ask it to generate a group picture And here's the first image it gave me and first I can see that it is not generated in the aspect ratio of 16 by 9 And also it generated two Lucky, which is actually pretty normal when DALLE generates images of multiple characters. But I'll show you how to correct this. I first would like to ask it to generate always in 16 by 9.

so you can see that even if in the instructions sometimes it doesn't read everything in there. technically the things that you put in the instructions after you save it should always take that into consideration. But sometimes it just doesn't.

I don't know why. That's why you need to come here to correct it sometimes and to make sure everything is good. So it says that is it has updated it.

I want to check by going to configure and instruction and see what it changed. after I updated the 16 by 9 ratio it removed the base prompt So I will have to add it back. so always check back to the instructions to see if it has removed any important thing before you save.

that's why I also said to have a separate document to include all the details So it doesn't do anything like this and you lose the details of your other information. This copy. My computer is super slow uuuurgh Okay.

And read through it and make sure everything is good and close. go back to create and let's test again give me image of the three in the woods. Woah, it's a really weird come to Christmas town I probably accidentally clicked this so it generated an image on the right.

So it does give me a sixteen by nine now. But as you can see, the details of the characters are not all correct. For example, the shirt should be all plain yellow.

And also Yoko should have a white shirt like like this one. But something like this or even like this is easily correctable and I will teach you how to do it. So let's download this and this and let's save.

and publish to only me and confirm. the first image I asked it to generate gave me two dogs, but this is actually correctable. I will show you how to do it later And then I test to see if it can create something with more action.

Like chasing a train. The first image it gave me is not really the style I'm looking for, so I'm going to ignore it. And the second one, somehow it generated a a forth character.

And then Marcus doesn't really look like him on the side. Yes. Sometimes like something like this happen.

If it's too hard to correct, well, just ignore it and try again. I think this one is pretty similar to what I'm looking for. I really like this one.

And then I try to upload a reference image from Up and trying to ask it to generate the same thing the first time it gave me an error. But the second time it gave me something like this, I think it get the gist. I asked it to generate again.

And this time, the character of Marcus is not right. So I tried again. This one I don't like even worse.

and I tried to ask for the gen_I. D. This is the process.

You just have to try again and again. and then I upload another image from Up to recreate the same scene. It give me this one.

And the extra dog is easily correctable. And this weird detail, I try to use the image from Soul, as a reference I think it completely capture the emotion and expression from the reference image, and then I ask it to do one with Marcus. so overall I can say that this bot is definitely not perfect, but there are just a lot of things you can do with it, with the capabilities of uploading your reference images and assigning the details of each character and putting them all in the same environment.

With that said, I'm going to teach you how to correct the wrong details that it generated. the easiest way to correct the detail is to use Canva Plus. you first upload your design using the new design.

custom size. generate new design, Click on the image. Click Edit Photo.

The Magic studio would show on the left side. Click Magic Eraser. Adjust the brush size and simply just paint over the unwanted parts.

another thing I want to address on this image is I want Yoko to wear a skirt. go to magic edit so we can adjust the brush size and brush over the part that we don't want. and then hit continue.

Describe. I would put skirt in here and generate. it would then generate four results for you to choose from.

If you don't like any of them, just click “generate new results”. Select the one that you like and click done. To save, click Share and click Download choose the file type.

Adjust the size and other details. Download. And now you have an image that you can use.

I hope what I shared with you today is helpful for you. If you want to see how to turn these images into animations, watch the next video here and I will walk you through with a step by step process. We’ll see you there.