I love the method you used to get image consistency: construct a conversation history with the model where it explicitly accepts each image and agrees to use it:
User: Previous page prompt: [Previous Prompt 1]
User: [Image 1]
Assistant: I've received this image and will maintain visual consistency with it.
User: Previous page prompt: [Previous Prompt 2]
User: [Image 2]
Assistant: I've received this image and will maintain visual consistency with it.
User: Previous page prompt: [Previous Prompt 3]
User: [Image 3]
Assistant: I've received this image and will maintain visual consistency with it.
User: Please maintain character appearance, art style, and color palette consistency with these previous illustrations when creating the next image.
Assistant: I'll ensure the characters, art style, and colors remain consistent with the previous illustrations.
User: Generate an illustration for a children's ABC story: [Current Prompt]. Make it colorful, child-friendly, and in a consistent style with any previous images. Include diverse characters with different ethnicities, genders, and abilities. Ensure representation is natural and authentic.
I'm curious: did you settle on this after trial and error? I mean, did having the assistant explicitly agree increase consistency, compared with just asking nicely at the end?