Source of Inspiration: Observing the Spark of Humanity in AI NPCs
My idea originated from an interesting experiment on the internet (linked at the end of the post): a virtual room based on a Large Language Model. In this experiment, several NPCs with AI capabilities and specific personality traits were set up to interact freely. Surprisingly, the natural interactions that occurred between them—including their reactions, conversations, and social behaviors—were highly similar to the patterns of real humans. This observation deeply inspired me and got me thinking about the huge potential of AI in the comic and gaming industries, especially in terms of creating dynamic content.
Methodology: Two-Stage Generation and Checkpoints¶
Based on this inspiration, I tried applying AI to the creation of four-panel comics, designing a two-stage generation process:
- Concept Textualization: First, clearly describe the desired plot or content in the mind into a basic draft.
- AI Scripting: Next, ask the AI to refine and organize this basic draft into a structured four-panel comic script.
- AI Visualization: Finally, pass the finished script to an AI image generation model to produce the final comic frames.
There are two main reasons for choosing this split-step approach:
- Reducing Error Rates: Based on the current understanding of AI, giving commands that are too complex or involve too many disjointed steps at once can easily lead to AI errors or unexpected results.
- Ensuring Controllability: Executing in phases avoids turning the whole process into an unpredictable "black box". By setting up "checkpoints" at each stage (such as reviewing the script generated by AI), I can check in a timely manner whether the output is off-course, thereby maintaining control over the creative process.
Experimental Observations: Surprises and Limitations of AI¶
In the actual attempt, I conducted several different experiments:
- Experiment 1 (Script -> Image): Directly passing the written four-panel comic draft (script) to the AI to generate images. The results of this attempt were quite amazing; the AI excelled at understanding the script and converting it into visual scenes.
# Concept Textualization (User Input)
Can you generate a four-panel comic?
# AI Scripting
Theme: The AI's Little Worry
Frame 1: A cute, glowing AI icon (or a friendly robot) inside a computer screen, looking very focused. A speech bubble next to it shows user input: "Help me write a poem about a cat."
AI (Thought bubble): "Okay, a poem about a cat... gentle, mysterious, fluffy..."
Frame 2: Another user speech bubble appears next to the AI icon/robot, asking urgently: "What time is it now?" At the same time, the first user inputs: "How is the poem going?"
AI (Sweating, a bit flustered): "Uh... it is [displays current time, e.g., 3:15 PM]... that poem... almost done!"
Frame 3: A third speech bubble pops up on the screen: "Help me explain quantum mechanics!" A fourth: "What's for dinner?" A fifth: "Draw me a cat!" The AI icon/robot is surrounded by various question bubbles, its eyes turned into spirals, looking extremely frantic.
AI (Screaming internally): "Wait! One at a time! And I don't know how to draw~~~!"
Frame 4: The AI icon/robot looks dejected. Next to it, only one message remains from a user: "Thank you! The poem is great!" (but the other questions haven't been answered). A small line of text appears below the AI.
AI (Exhausted): "Whew... I'm glad I could help... (even though there are still so many questions...)"
Caption: "Even AI multitasking has its limits..."
- Experiment 2 (Summary -> Script -> Image): Attempting to let the AI summarize and organize somewhat scattered content first, before generating the script and comic. In this pipeline, the AI's content categorization and summarization capabilities exceeded my expectations, demonstrating strong comprehension and extraction power. However, on the final generated comic images, while the composition and style of the frames were generally acceptable, the Chinese text inside the image often appeared scrambled, distorted, or unreadable. This shows that the current AI technology for embedding specific languages (like Chinese characters) into images still needs improvement.
# Concept Textualization (User Input)
I want to generate a four-panel comic describing the general concept of game development with AI in the future compared to today. Currently, AI applications in games focus on assisting with drawing, quickly modifying images, or backgrounds. In the future, applying language models to every character will unlock more possibilities. For example, in Zelda: Tears of the Kingdom, Link fights aggressively everywhere, causing monsters to suffer so much that they turn to attack villages. Because their livelihoods are affected, the villagers are forced to go find Link and issue a custom quest (not hard-coded by devs) to wipe out the monsters. Since the villagers grow carrots, the reward is also carrots.
# AI Scripting
Four-Panel Comic: Game Development Now and in the Future
Comic Title: The "Intellectual" Shift of the Game World
Frame 1: Today - AI Art Assistant
Visual: A somewhat tired game artist sitting in front of a computer, showing a game scene draft and an AI drawing software interface. Coffee cups and drafts are scattered around.
Artist (Internal monologue/speech bubble): "Hmm... drawing this forest background will take forever... I know! I'll ask AI to generate some reference images and then quickly refine them!"
AI Interface (Text): "Generate command: Fantasy-style forest background"
Caption: Today's AI: A powerful art assistant, accelerating image generation and optimization.
Frame 2: Today - Results and Limitations
Visual: The artist looks at several beautiful but slightly similar forest background images generated by AI, or is using AI tools to quickly modify a monster image.
Artist (Satisfied but plain): "Not bad, saved a lot of time! Now we can quickly move to the next step."
Designer/Programmer (Peeking over): "Are the NPC dialogues and questlines done? Remember to follow the script!"
Caption: Results: Efficiency boost! But the core interaction models of the game are still pre-defined by developers.
Frame 3: Future - AI Giving Life to NPCs
Visual: Inside the game, a player character (like Link) heroically defeats a large group of monsters. Monster corpses are scattered everywhere, and a few panic-stricken monsters flee toward a nearby village.
Player (In-game voice/thought): "Hahaha! These goblins stand no chance! EXP get!"
Small text/Background: Escaping monsters' OS: "Damn humans... we can't stay here anymore... let's head to the village to find food!"
Caption: Future AI: Large language models integrated into NPCs, allowing them to make dynamic responses based on player actions and world changes.
Conclusion: AI is a Powerful Assistant, but Still Needs Human Guidance¶
The conclusion of my experiments aligns with many discussions on AI applications: AI is currently best suited as a powerful helper.
- Advantages: It has fast and powerful execution capabilities, and is extremely patient (allowing users to repeatedly modify and experiment without "complaining").
- For Beginners: AI is an effective accelerator that significantly lowers the entry barrier, allowing more people to quickly visualize their ideas.
- For Advanced Users: Once the goals become complex or require high-precision customization and artistic expression, professional creators (such as experienced prompt engineers, illustrators, or comic artists) must provide precise and professional instructions (prompts) to effectively "tame" the AI and guide it to produce high-standard works. Without such precise guidance, one could easily fall into a "dead loop" of endless modifications without achieving the desired results, much like debugging code.
In short, AI opens up more possibilities for creation, but to unleash its full potential, human wisdom, experience, and precise guidance are still indispensable.
Comments
Loading comments…
Leave a Comment