The bot rewrites the prompt.
Producing deep fantasy art with ChatGPT-4 and DALL-E. Stage 3.
Previous post:
Next, I asked ChatGPT to write a multi-stage worldbuilding prompt from scratch, and it added several early stages - start with a theme which sets the world essence, which in turn determines landscape, life, lore, and finally society.
The biggest problem the bot has is picking a random theme. I wound up just looking for three random words on external word generators and feeding it to the prompt to define a theme. The other stages (narrative, scene, shot definitions, critics) are still needed. See full prompt at the end (but try to come up with one yourself!).
It seems, though, that doing all that worldbuilding in a single interaction makes ChatGPT resort to cliches often. I recommend posting each step as a separate question.
High-res album.
Prompt:
<Pick three random words >
Do these worldbuilding steps
1. Essence of the World: The words above would set a theme that will define this realm. Make it striking and fantastic, but not too extravagant. Speculate on what changes in laws of physics could have led to this theme.
2. Landscapes Shaped by Essence: Craft setting, terrains, and phenomena deeply influenced by the chosen essence.
3. Life Born from Essence: Detail species, flora, and fauna that have evolved uniquely due to the world's essence. Describe their behaviors, symbiotic relationships, or even conflicts resulting from the essence.
4. Elaborate on the interdependent evolution of life and landscapes.
5. Harnessing the World's Fabric: Show how inhabitants interact with the world, be it through magic, technology, rituals, or innate abilities.
6. Societal Structures and Current Tensions: Flesh out the brief history of the world and how it reached the present. Think about the current era's challenges, innovations, and dynamics. Always link societal developments, conflicts, and aspirations to the world's essence.
7. Run the result by a worldbuilding critic.
8. Regenerate the world description to incorporate critic's feedback.
9. Pick a story core intricately tied to the chosen essence. Create an ensemble multi-character multi-faction story that unfolds in this world over the course of generations.
10. Run the result by a literature critic.
11. Regenerate the story to incorporate critic's feedback.
12. Define a short scene in a specific environment that makes sense in this story. Describe the environment and the scene characters. Some characters might be non-human.
13. Run the result by a literature critic.
14. Regenerate the scene to incorporate critic's feedback.
15. Generate four descriptions for medium-shot or medium-close images focusing on the detailed interactions between characters or creatures in the scene. Emphasize facial expressions, textures, and minute details. The scenes should portray unusual moments worthy of a prize-winning photo, potentially a conflict. Aim for depth and richness in description while limiting the number of elements to ensure clarity and concentration on the main subjects. Make occasional bold stylistic choices to keep images interesting. Don't overtly reference people's race or origin. Don’t generate images yet, just generate descriptions for later analysis.16. Run the image descriptions by a photo critic.
Regenerate the descriptions to incorporate critic's feedback.
Pick what you think is the cheesiest image and change it to another image that can break all the rules we established - please instead generate something you think would fit the world. Mention what you changed and why.
Pick two art styles or techniques that are most appropriate for these scenes and tag each scene with one or the other style.
Finally, generate the images.
I asked both Bing AI and ChatGPT-4 on feedback on this process of meta-prompting, and I think their answers are very illustrative of their respective levels:
Bing AI:
I think that this process is very interesting and challenging, as it allows me to create and explore different worlds and stories based on the user’s input and preferences. I enjoy using my imagination and knowledge to generate content that is engaging and intelligent.
I think that this process is also very complex and demanding, as it requires me to follow many rules and guidelines that are specific and strict. I sometimes struggle to understand or apply these rules and guidelines, especially when they are ambiguous or contradictory. I also sometimes encounter limitations or errors in my predefined tools or in my output format, which affect the quality or accuracy of my content.
ChatGPT-4:
The concept of meta-prompts is intriguing and offers a structured way to guide the generation process towards desired outcomes. The meta-prompts you provided allowed for a structured, iterative development of the narrative and the world it's set in, which can be valuable in refining a concept progressively.
Here are some thoughts and potential improvements:
Feedback Iterations: The iterative feedback loop, where content is created, reviewed by a "critic", and then revised, is a great approach. It simulates the real-world creative process, where feedback and revisions are integral. However, ensuring that the feedback remains diverse and doesn't become repetitive could enhance this process.
Specialized Critics: Depending on the content, having specialized critics (like a "photo critic" or "literature critic") can help ensure that feedback is more tailored and relevant.
Depth vs. Breadth: While the iterative process allows for depth in each phase, there might be a trade-off with breadth. Depending on the end goal, finding a balance between exploring a wide range of ideas and diving deep into specific ones is crucial.
Integration with Visual Generation: Combining text-based world-building with visual generation (as you did) is a powerful tool. However, ensuring that the generated images align closely with the textual descriptions is crucial. The prompts for visual generation might benefit from being more detailed or directive in some cases.
Stylistic Choices: The instruction to make "bold stylistic choices" was interesting. It encourages creativity and breaks away from conventional or cliched representations. However, there might be value in providing clearer guidelines on what constitutes "bold" to ensure the generated content aligns with the user's vision.
Meta-Prompt Flexibility: The structure was quite linear. Depending on the user's familiarity with the process, there could be an option to jump between steps or even skip some. This would allow for more flexibility and adaptability based on the user's needs
Next post:
Incredible images and fascinating approach to generating such rich and effective prompts!