I just learned that neither the Gen_IDs nor the seed numbers are stored across sessions, so including the GEN_ID may not influence the results. 😅What do you think?
Amazing! Been working on an ecommerce project using AI artwork, but was struggling so hard to get consistent characters. I made a few modifications to tailor my specific needs, and it turned out just great. You really made my day a lot better, Mia. Thank you so much ☺ Subscribed ! PS: Do you have a Discord channel where people can follow up and share ideas?
So glad it was helpful and thanks for subscribing! You've made my day! I don't have a Discord channel right now but it is a great idea and I may get one in the future. 😄
Thank you for your great tutorials. Unfortunately, I have spent so much time and money trying to get consistency, I may just have to give up. I simply cannot afford to add Adobe too. I just started using Dall E 4 but I’m not getting my hopes up. Thanks, again!
That's a good question. You can try enter a base prompt and input reference images for each character in the same custom GPT bot just like what I did on one character. I would also share the relationships of these characters to the GPT and add more instructions to avoid the mix-up. When there are more characters, there is for sure more chance for AI hallucinations to happen. I am currently testing on this and plan to share my findings in a more detailed video once I have some conclusive insights 🙂
Exactly like @iammiameow pointed out. But, I still noticed that Dall_E 3 tends to mix up character features or come up with something completely different than I asked it to do (AI hallucinations) if there are conflicting sentences in the instructions, so make sure you check that out first. Then I experimented with something that gave me some interesting results in consistency or at least minimize substantially the number of regens. I started out the conversation with an "ignition prompt" (yeah, I just came up with that one, lol). It is the same prompt I used in the instructions when building the GPT, but with a few mods. For example, this is my ignition prompt: "Use this prompt VERBATIM. Create an anime-style scene depicting two characters exploring Alabama and its iconic places. The first character is an Anime male in he's 30s named Silva and has short black hair and brown eyes and also a bit fit, wearing casual vacation clothing. The second character is an Anime woman in her 30s named Lora (slim; fit; mid-size chest), has long orange hair and blue eyes, dressed in bright, comfortable travel clothes. Make the characters' expressions joyful and lively, capturing the spirit of adventure and the beauty of Alabama. This prompt sets the scene for a series of images, with your characters consistently portrayed across each image generation, enjoying various iconic locations and activities." (Do not include the Gen_ID in this prompt, but maintain the original one inside GPT's builder instructions. This will help a great deal). If the first one comes up great, use it straight away. If not, then generate more images using the edit button until you find the one you want to start out as a baseline. What I speculate this does is it iterates your prompt verbatim, and then cross-references the instructions and images inside the GPT, outputting a result very similar to what you have imagined. Quick note: Be mindful when describing body features. Dall-E 3 tends to censure things it thinks are NSFW resulting in generation errors. Now is the time to ask GPT for the Gen_ID of the image you chose as a baseline after the ignition prompt. For example: You: "what is the Gen_ID of this image?" Your GPT will respond: "The Gen_ID for the image created is 4zjH0X0vvoaNqELe" Great! We now have two guaranteed Gen_ID's for Dall-E 3 to reference and reinforce the quality of your outputs. The one in the builder instructions, and the one you will be using in the current chat window. So now, I prompted again with: "Create an anime-style scene depicting two characters exploring the streets in the city of Mobile in Alabama as the background. Remember, the first character is an Anime male named Silva in he's 30's and has short black hair and brown eyes and also a bit fit, wearing casual vacation attire. The second character is an Anime woman named Lora (slim; fit; mid-size chest), has long orange hair and blue eyes, also dressed in bright, comfortable travel clothes. Make the characters' expressions joyful and lively, capturing the spirit of adventure and the beauty of Alabama. Use Gen_ID:4zjH0X0vvoaNqELe" Notice that now I have modified slightly the ignition prompt, changing only the specific parts like the location I want my characters to be in, and at the end I include the Gen_ID from of the image GPT generated in this chat window. Now the fun part begins. This might sound silly or crazy, however, it does make sense. GPT is a natural language model, so it would make sense to use, well, natural language. Also, and again, it might sound silly, but treat GPT kindly. I noticed that a great deal of generating consistent characters is due to me complimenting its great results. So, from here on out, just start by having a natural conversation with short and simple sentences. For example: "Very good! Now let's take them on a walk in the historic Sloss Furnaces in Birmingham, Alabama at sunset. Maintain consistency. Use Gen_ID: muAbRROFu2lyFd5w as reference for the characters." (Always include this at the end). And then: "Nice job! Now, they are excited to go see the rockets flying at the Huntsville Space and Rocket Center in Alabama. Let's take them there shall we? Maintain consistency. Use Gen_ID: muAbRROFu2lyFd5w as reference for the characters." ...and just keep doing the same thing over and over again. Most of the time, the first image is bang on the money. Yes, there will be the occasional error or lack of consistency. If that happens, try to regen or move them to another location all together. I would love to share screenshots of this experiment with multiple characters (well, at least two) in one scene\image. If there is anyway to do that, please let me know guys.
Wow, these are some great findings. I will give it a shot. I love the part that you compliment GPT's results and see better outcomes. It kind of makes sense that a natural language model would benefit from positive reinforcement when adjusting to favorable outcomes.Thanks for your insights!
I have this link, which was included in the description: mia-meow.notion.site/Story-Illustrator-GPT-Notes-c029fc6a7a6e4399b248459080ce5ef5?pvs=4 Hope it can help :)
I just learned that neither the Gen_IDs nor the seed numbers are stored across sessions, so including the GEN_ID may not influence the results. 😅What do you think?
Thank you so much for your video. I already make Bot follow your instructions.🙊🙊
So glad that it helps!! 😀
Amazing! Been working on an ecommerce project using AI artwork, but was struggling so hard to get consistent characters. I made a few modifications to tailor my specific needs, and it turned out just great. You really made my day a lot better, Mia. Thank you so much ☺ Subscribed ! PS: Do you have a Discord channel where people can follow up and share ideas?
So glad it was helpful and thanks for subscribing! You've made my day! I don't have a Discord channel right now but it is a great idea and I may get one in the future. 😄
@@miameowai Sure thing 😁
Thank you for your great tutorials. Unfortunately, I have spent so much time and money trying to get consistency, I may just have to give up. I simply cannot afford to add Adobe too. I just started using Dall E 4 but I’m not getting my hopes up. Thanks, again!
Yes, it is tough. You can try Leonardo, it is free and they just added character consistency feature.
Yes, thanks!
enjoyed video keen to try something similar would it be possible to get the prompts that you used so i can modify it to my needs
I am glad that you have enjoyed the video! Thanks for watching! Here is the link to the notion notes and prompts I used for this GPT. rb.gy/zvmea6 😀
How do you make it work if you wanted to create n interactive scene photo with more than one consistent character, say a book.
That's a good question. You can try enter a base prompt and input reference images for each character in the same custom GPT bot just like what I did on one character. I would also share the relationships of these characters to the GPT and add more instructions to avoid the mix-up. When there are more characters, there is for sure more chance for AI hallucinations to happen. I am currently testing on this and plan to share my findings in a more detailed video once I have some conclusive insights 🙂
Exactly like @iammiameow pointed out. But, I still noticed that Dall_E 3 tends to mix up character features or come up with something completely different than I asked it to do (AI hallucinations) if there are conflicting sentences in the instructions, so make sure you check that out first. Then I experimented with something that gave me some interesting results in consistency or at least minimize substantially the number of regens. I started out the conversation with an "ignition prompt" (yeah, I just came up with that one, lol). It is the same prompt I used in the instructions when building the GPT, but with a few mods. For example, this is my ignition prompt:
"Use this prompt VERBATIM. Create an anime-style scene depicting two characters exploring Alabama and its iconic places. The first character is an Anime male in he's 30s named Silva and has short black hair and brown eyes and also a bit fit, wearing casual vacation clothing. The second character is an Anime woman in her 30s named Lora (slim; fit; mid-size chest), has long orange hair and blue eyes, dressed in bright, comfortable travel clothes. Make the characters' expressions joyful and lively, capturing the spirit of adventure and the beauty of Alabama. This prompt sets the scene for a series of images, with your characters consistently portrayed across each image generation, enjoying various iconic locations and activities." (Do not include the Gen_ID in this prompt, but maintain the original one inside GPT's builder instructions. This will help a great deal).
If the first one comes up great, use it straight away. If not, then generate more images using the edit button until you find the one you want to start out as a baseline. What I speculate this does is it iterates your prompt verbatim, and then cross-references the instructions and images inside the GPT, outputting a result very similar to what you have imagined.
Quick note: Be mindful when describing body features. Dall-E 3 tends to censure things it thinks are NSFW resulting in generation errors.
Now is the time to ask GPT for the Gen_ID of the image you chose as a baseline after the ignition prompt. For example:
You: "what is the Gen_ID of this image?"
Your GPT will respond: "The Gen_ID for the image created is 4zjH0X0vvoaNqELe"
Great! We now have two guaranteed Gen_ID's for Dall-E 3 to reference and reinforce the quality of your outputs. The one in the builder instructions, and the one you will be using in the current chat window.
So now, I prompted again with:
"Create an anime-style scene depicting two characters exploring the streets in the city of Mobile in Alabama as the background. Remember, the first character is an Anime male named Silva in he's 30's and has short black hair and brown eyes and also a bit fit, wearing casual vacation attire. The second character is an Anime woman named Lora (slim; fit; mid-size chest), has long orange hair and blue eyes, also dressed in bright, comfortable travel clothes. Make the characters' expressions joyful and lively, capturing the spirit of adventure and the beauty of Alabama.
Use Gen_ID:4zjH0X0vvoaNqELe"
Notice that now I have modified slightly the ignition prompt, changing only the specific parts like the location I want my characters to be in, and at the end I include the Gen_ID from of the image GPT generated in this chat window.
Now the fun part begins. This might sound silly or crazy, however, it does make sense. GPT is a natural language model, so it would make sense to use, well, natural language. Also, and again, it might sound silly, but treat GPT kindly. I noticed that a great deal of generating consistent characters is due to me complimenting its great results. So, from here on out, just start by having a natural conversation with short and simple sentences. For example:
"Very good! Now let's take them on a walk in the historic Sloss Furnaces in Birmingham, Alabama at sunset.
Maintain consistency. Use Gen_ID: muAbRROFu2lyFd5w as reference for the characters." (Always include this at the end).
And then:
"Nice job! Now, they are excited to go see the rockets flying at the Huntsville Space and Rocket Center in Alabama. Let's take them there shall we?
Maintain consistency. Use Gen_ID: muAbRROFu2lyFd5w as reference for the characters."
...and just keep doing the same thing over and over again.
Most of the time, the first image is bang on the money. Yes, there will be the occasional error or lack of consistency. If that happens, try to regen or move them to another location all together.
I would love to share screenshots of this experiment with multiple characters (well, at least two) in one scene\image. If there is anyway to do that, please let me know guys.
Wow, these are some great findings. I will give it a shot. I love the part that you compliment GPT's results and see better outcomes. It kind of makes sense that a natural language model would benefit from positive reinforcement when adjusting to favorable outcomes.Thanks for your insights!
@@miameowai Great, Mia! Let us know how it turned out. 😉
Also I need your full process of creating an illustration book plz .
I will note this down. Thanks for the suggestion :)
Can you please tell me how we can get access of chat gpt to generate images is there any subscription or any other thing to do please guide one by one
Yeah, you will have to use the paid version for image generation at this time. Or you can use bing’s image creator for free.
omg... digital art production is dead ... a brilliant video, thank you.
I wouldn’t say it’s dead, I think it just went to another dimension lol, thanks for watching!
Can you share your prompts as pdf . Pleaaaaase . To make our life easier
I have this link, which was included in the description: mia-meow.notion.site/Story-Illustrator-GPT-Notes-c029fc6a7a6e4399b248459080ce5ef5?pvs=4
Hope it can help :)