Objective

This lab introduces students to generative AI in the visual domain through the use of Stable Diffusion, a powerful text-to-image model. Students will interact with a WebUI hosted via Docker on the Jetson Orin Nano and learn how prompt engineering and hyperparameter tuning affect the quality and diversity of generated images.

Learning Outcomes

• Understand the architecture and operational pipeline of Stable Diffusion.
• Generate images from text prompts using various sampling methods and model configurations.
• Analyze the effect of parameters such as sampling method, CFG scale, batch size, and seed value.
• Develop prompt engineering skills to influence image quality and structure.

Lab Tasks

1. Environment Setup: Pull and launch the Docker container dustynv/stable-diffusion-webui:r36.2.0. The UI is hosted locally on http://0.0.0.0:7860.
2. Basic Prompt Execution: Start with a fixed prompt (e.g., \An orange cat wearing a cowboy hat") and generate base output.
3. Hyperparameter Tuning:
   • Experiment with 3 different sampling methods and analyze qualitative differences.
   • Vary sampling steps (e.g., 10, 50, 150) and observe changes in resolution and detail.
   • Adjust CFG Scale to influence prompt adherence.
   • Test different batch count and size combinations to study their impact on diversity.
   • Use different seed values and compare outputs to evaluate randomness and reproducibility.
4. Documentation: Capture and annotate screenshots to record parameter effects and insights.

Expected Deliverables

• Image outputs from all test configurations.
• Short report on Stable Diffusion architecture and prompt-performance relationships.
• Summary of findings from hyperparameter variations.

Objective

This lab explores the application of Large Language Models (LLMs) for real-time, prompt-based natural language generation. Students will run a lightweight LLM—TinyLLaMA—within a WebUI hosted via Docker, and investigate the role of different hyperparameters in shaping output quality and coherence.

Learning Outcomes

• Understand the workflow for deploying and interacting with LLMs locally on edge devices.
• Learn about decoding strategies and hyperparameters that control generation style, fluency, and creativity
• Evaluate and compare outputs across different parameter settings.

Lab Tasks

1. Model Setup: Pull and launch the Docker container dustynv/text-generation-webui:r36.2.0. Use the terminal-generated local URL to open the WebUI.
2. Model Download and Load: Use the WebUI interface to download the model TheBloke/TinyLlama-1.1B-Chat-v1.0-GGUF and load the quantized model file tinyllama-1.1b-chat-v1.0.Q3_K_M.gguf.
3. Prompt Execution: Start with a base prompt (e.g., \Tell me what you know about New Jersey Institute of Technology") and run inference.
4. Parameter Exploration: Vary at least 3 values for each of the 10 hyperparameters available in the WebUI (e.g., temperature, top-p, top-k, repetition penalty).
5. Output Analysis: Record changes in response length, creativity, hallucination, and repetition under different parameter settings.

Expected Deliverables

• Screenshots of outputs for each parameter setting (minimum 30 examples).
• Concise write-up on the TinyLLaMA architecture and inference control strategies.
• Comparative analysis of hyperparameter effects on generation quality.