AI Lab 06: Visual Generative AI & The Character Consistency Challenge
- Estimated Time: 5 hours
- Tools Required: Leonardo.ai (Recommended for advanced controls) OR Microsoft Copilot Designer. Both are free to use.
- Hardware Required: A computer with an internet connection.
Lab Objectives
- Understand how Text-to-Image models (Diffusion Models) interpret prompts differently than Text-to-Text LLMs.
- Master the "Anatomy of an Image Prompt" (Subject, Environment, Lighting, Camera, Style).
- Core Task: Experience the difficulty of maintaining "Character Consistency." You will attempt to generate the exact same original character performing different actions in different scenes.
- Final Project Kickoff: Use these tools to generate the visual concept art for your group's final project.
5-Hour Schedule & Step-by-Step Breakdown
Phase 1: Platform Setup & The Anatomy of a Prompt (0.5 Hour)
Phase 1: Platform Setup & The Anatomy of a Prompt (0.5 Hour)
- Go to Leonardo.ai (or Copilot Designer) and create a free account.
- In Leonardo.ai, click on Image Generation. Familiarize yourself with the interface: the prompt box, negative prompt box, aspect ratio settings, and the generation button (which costs credits/tokens).
- Understand Image Prompts vs. Text Prompts:
- LLMs (like ChatGPT) want conversational instructions: "Please write me a story about a dog."
- Image AI (Diffusion models) want comma-separated visual tags: "A golden retriever, sitting in a sunny park, cinematic lighting, shot on 35mm lens, photorealistic, 8k."
Phase 2: Image Generation Experiments (2.5 Hours)
Phase 2: Image Generation Experiments (2.5 Hours)
In this phase, you will learn to precisely control the AI's "brush" and then attempt to lock down a character's identity.
Experiment 1: The Prompt Stacking Technique (1 Hour)
Your goal is to see how adding specific modifiers drastically changes the output.
- Base Prompt: Generate a simple prompt: "A futuristic city." (Save the image).
- Add Style & Medium: Add tags to change the art style. Generate: "A futuristic city, cyberpunk style, neon colors, digital painting, ArtStation."
- Add Camera & Lighting: Now, control the virtual camera. Generate: "A futuristic city, cyberpunk style, neon colors, heavy rain, glowing fog, cinematic lighting, dramatic shadows, wide angle lens, bird's-eye view."
- Observe: Compare the three images. Notice how adding specific photography and lighting terms transforms a generic image into a professional-looking concept art piece. (Take a screenshot of your final, fully stacked prompt and image).
Experiment 2: The Character Consistency Challenge (1.5 Hours)
AI naturally wants to generate a different face every time you click "Generate." Your challenge is to make the AI remember who your character is.
- Create the Base Character: Write a highly detailed description of a unique character. Do not use famous people (like "Tom Cruise").
- Example: "A 30-year-old female detective, wearing a yellow trench coat, short blue hair, round glasses, steampunk style."
- Generate it until you find a face you really like. Save this image.
- The Challenge: Now, you must generate this exact same woman doing three different things:
- Action 1: Drinking coffee in a cafe.
- Action 2: Running down an alleyway.
- Action 3: A close-up portrait looking angry.
- Techniques to Try:
- Option A (Prompt Locking): Keep the exact visual description of her face and clothes in every single prompt, just changing the action at the end.
- Option B (Image Reference/Character Reference): In Leonardo.ai, upload your saved Base Character image into the "Image Guidance" or "Character Reference" tab to force the AI to copy her face.
- Observe: Did her clothes change? Did her hair get longer? Did her facial features morph? (Save your best attempts for the journal).
Phase 3: Final Project Kickoff - Concept Art (0.5 Hour)
Phase 3: Final Project Kickoff - Concept Art (0.5 Hour)
- Communicate with your group members. Decide on the theme/product/story for your Week 8 Final Project.
- Use the AI tools from this lab to generate 2 to 3 pieces of Concept Art for your project. (e.g., If you are designing an AI virtual assistant app, generate what the mascot looks like, or what the app interface might look like).
Phase 4: Writing the AI Lab Journal (1.5 Hours)
Phase 4: Writing the "AI Lab Journal" (1.5 Hours)
Complete your "AI Lab Journal" in a Word or Markdown file.
AI Lab Journal Report
AI Lab Journal: Generative AI & Character Consistency
1. Lab Setup & Key Definitions
- Platform Used: [e.g., Leonardo.ai]
- What is a "Diffusion Model"? Explain in 1–2 sentences how AI generates images from "noise".
2. Prompt Stacking Evolution Table
| Stage | Prompt Used | Visual Observation (What changed?) |
|---|---|---|
| 1. Base | A futuristic city. | Generic, boring cityscape. |
| 2. Style | [Your prompt with style tags] | [Describe how the mood changed] |
| 3. Camera/Light | [Your prompt with camera/lighting tags] | [Describe the impact of camera angles/lighting] |
3. Test Results & Screenshot Evidence
- Experiment 1 (Prompt Stacking):
[Insert a screenshot of your final, fully stacked futuristic city image here]
- Experiment 2 (Character Consistency): How hard was it to keep the character looking identical? What specifically kept changing—her face, clothes, or hair?
[Insert the screenshot of your Base Character][Insert the screenshot of Action 1 (Drinking coffee)][Insert the screenshot of Action 2 (Running)]
4. Group Project Visuals (Kickoff)
- Our Final Project Idea: [1–2 sentences explaining what your group is building]
[Insert 1–2 pieces of AI-generated concept art for your final project here]
5. Core Reflections
- Question 1: Why do you think Diffusion Models struggle so much with "Character Consistency" compared to 3D rendering software?
- Question 2: If you use an AI tool to generate an award-winning piece of art using a prompt you spent 5 hours perfecting, do you think you should own the copyright to that image? Why or why not?
🎯 Pro-Tips for Students (How to Avoid Common Pitfalls)
- The "Prompt Soup" Mistake: Don't just throw 100 random adjectives into the prompt box (e.g., "beautiful, awesome, masterpiece, best quality, trending on artstation"). Be deliberate. Structure your prompt logically: [Subject] + [Environment/Action] + [Lighting] + [Camera/Medium].
- Negative Prompts are your Friend: If your character keeps generating with deformed hands or the wrong hat, use the "Negative Prompt" box. Typing "hat, extra fingers, blurry, distorted" tells the AI specifically what to subtract from the noise.