Generative AI Course Module Catalog: A Complete Path from Intro to Practice ​

We · October 2026

This article is the second in the "AI Course Design Guide" series. The first article discussed the overall design framework of an AI curriculum. This one drills down one layer — what specific modules should a generative AI course include? What does each module teach, what tools does it require, and what are the environment requirements?

If the three-layer framework from the first article is the "skeleton," this article is the "flesh and blood."

Many schools and training institutions want to launch generative AI courses but face a practical question: what exactly should be taught?

Borrow a university computer science AI course directly — too theoretical, non-majors won't understand it. Only teach how to use ChatGPT — too shallow, students finish feeling like they "know a bit of everything but nothing in particular." The courses on the market are even more varied — some are called "AI Application Engineer," some "Large Model Development in Practice," some "Generative AIGC Master Class." The names get ever grander, but what the content actually differs on is hard to articulate.

We compiled mainstream generative AI course systems on the market — including the layered architecture of corporate internal training, the complete chain of bootcamps, and the module design of vocational training — and, combined with the environment requirements from teaching practice, distilled a modular generative AI course catalog.

This catalog isn't for any one specific course. It's a "module library" — schools, training institutions, and corporate training departments can pick modules from it based on their target audience and available hours, and combine them into a course. Each module includes: what to learn, what learners can do after completing it, what tools are required, and what the local deployment requirements are.

Three Tracks, Pick as Needed ​

Generative AI courses can be divided into three broad tracks by audience and depth. These are not a linear progression from low to high — they are different learning paths. Some people only need to reach "Application"; some need to reach "Development."

Specific modules are listed by track below.

I. Literacy Track: AI Literacy Modules ​

This is the most foundational layer, aimed at everyone. The goal is not to train AI engineers, but to give everyone basic AI literacy — knowing what AI can and cannot do, using AI tools to boost efficiency, and understanding AI's boundaries and risks.

Module 1: AI Fundamentals ​

What to learn:

• What is generative AI? How does it differ from traditional AI?
• Large models, diffusion models, multimodality — core concepts in plain language
• Key milestones in AI's development
• Overview of mainstream AI products and services (conversation, image, video, speech, code)

What learners can do after completion:

• Understand AI-related news and discussions
• Know which types of AI tools suit which scenarios
• Avoid being swayed by either "AI omnipotence" or "AI uselessness"

Required tools:

• Online AI services (for demonstration; no local deployment required)

Local deployment requirements: None

Module 2: Prompt Engineering Fundamentals ​

What to learn:

• What does a good prompt look like?
• Structured prompting (role setup, task description, output format, constraints)
• Context management and conversation techniques
• Common prompt pitfalls and optimization methods
• Prompt evaluation and iteration

What learners can do after completion:

• Use AI to efficiently handle writing, translation, summarization, brainstorming, and similar tasks
• Write clear, effective prompts that yield stable-quality outputs
• Know when AI answers are unreliable and how to verify them

Required tools:

• Conversational AI tools

Local deployment requirements:

• Online tools can be used for teaching
• For hands-on local practice: Ollama + a 7B-class quantized model (around 6-8GB VRAM)

Module 3: AI Ethics and Compliance ​

What to learn:

• Copyright issues of AI-generated content
• Data privacy and personal information protection
• Deepfakes and content authenticity
• Algorithmic bias and fairness
• AI regulatory trends across countries and regions
• Principles for responsible AI use

What learners can do after completion:

• Use AI tools compliantly in their work
• Identify risks in AI-generated content
• Know which data should not be casually fed to AI

Required tools: None (discussion-based)

Local deployment requirements: None

Module 4: AI and Work Efficiency ​

What to learn:

• AI-assisted document writing and processing
• AI-assisted data analysis and spreadsheet handling
• AI-assisted information retrieval and organization
• AI-assisted learning and knowledge management
• How to embed AI tools into workflows

What learners can do after completion:

• Use AI to boost efficiency in everyday work
• Design AI-assisted workflows for specific work scenarios
• Know which tasks are suitable to delegate to AI and which are not

Required tools:

• Conversational AI + AI plugins for office software

Local deployment requirements:

• Online tools can be used at the literacy stage
• For advanced use, a locally deployed large model can process sensitive documents

II. Application Track: AI Tools and Creation Modules ​

This layer targets practitioners who need to produce concrete outcomes with AI — designers, content creators, operations staff, product managers, and the like. The core is "make things with AI tools." Learners don't need to understand underlying principles, but they must master the tools and workflows.

We note: generative AI models iterate quickly. The specific models mentioned here are only representative options at the current stage. Course design should center on principles and methods (diffusion model principles, workflow design, prompt methodology, etc.). Specific model choices can be refreshed each semester by referencing the latest rankings on platforms like Hugging Face and Civitai. Tools and methodology are stable; models are fluid.

Module 5: Image Generation — Fundamentals and Advanced ​

What to learn:

• Basic principles of diffusion models (at the concept level; no math required)
• Use of text-to-image, image-to-image, and inpainting
• Prompt techniques (positive/negative, weights, style description)
• ControlNet control (line art, depth, pose, segmentation, etc.)
• Concept and use of LoRA models
• Character consistency solutions such as IP-Adapter and InstantID
• Image post-processing and optimization

What learners can do after completion:

• Use AI to generate high-quality illustrations, posters, and concept art
• Use ControlNet to control composition and structure
• Use LoRA to generate specific styles or specific characters
• Complete an AI image creation workflow from concept to finished piece

Required tools:

• ComfyUI (core tool)
• Image generation models (FLUX.2 series, Qwen-Image, etc.)
• ControlNet model sets
• LoRA models (style, character)

Local deployment requirements:

• GPU: RTX 3060 12GB minimum; RTX 4070/5070 12GB+ recommended; with ample budget, RTX 5080/5090 or professional GPUs
• Model files: 30-80GB total (base models + ControlNet + several LoRAs)
• Dependencies: Python, CUDA, PyTorch, ComfyUI and custom nodes
• Difficulty: Medium — many dependencies; complex model management

Module 6: Video Generation and AI Editing ​

What to learn:

• Basic principles of text-to-video and image-to-video
• Characteristics and suitable scenarios of mainstream video generation models
• Prompt techniques for video generation
• Multi-shot video generation and stitching
• AI-assisted editing (auto-cut, transitions, subtitles)
• Video style transfer and restoration

What learners can do after completion:

• Use AI to generate short video clips
• Design multi-shot AI video workflows
• Use AI tools to boost video editing efficiency
• Complete an AI-assisted video production from script to finished cut

Required tools:

• Video generation models (MiniMax-H3, LTX 2 series, Wan 2.2, etc.)
• ComfyUI + video generation nodes
• AI editing tools
• Speech synthesis tools (for voiceover)

Local deployment requirements:

• GPU: RTX 4080/5080 16GB minimum; RTX 4090/5090 24GB+ recommended; labs can configure workstations or server GPUs
• Model files: 10-30GB per video model
• Disk: 160GB+ of available space
• Difficulty: High — video models demand more from hardware and environment

Module 7: AI Speech Synthesis and Processing ​

What to learn:

• Basic principles of TTS (text-to-speech)
• Characteristics and audio quality comparison of mainstream TTS models
• Principles and use of voice cloning
• Speech style control (emotion, speed, intonation)
• Multilingual speech generation
• Speech post-processing (denoising, restoration, mixing)

What learners can do after completion:

• Use AI to generate high-quality voiceover and narration
• Clone a specific voice and generate personalized speech
• Adjust emotion and style of speech
• Produce audio content such as video dubbing, audiobooks, and podcasts

Required tools:

• TTS models (IndexTTS, GPT-SoVITS, CosyVoice, etc.)
• Voice cloning tools
• Audio processing software

Local deployment requirements:

• GPU: RTX 3060 12GB is sufficient
• Model files: a few hundred MB to several GB
• Difficulty: Medium — some TTS tools are complex to install

Module 8: RAG Applications Primer (for non-developers) ​

What to learn:

• What is RAG? Why is it needed?
• The basic RAG workflow (document → chunking → embedding → retrieval → generation)
• Building and managing knowledge bases
• Methods for optimizing retrieval effectiveness
• Typical RAG scenarios (smart customer service, document Q&A, knowledge management)
• Building RAG applications with low-code tools

What learners can do after completion:

• Understand the principles and value of RAG
• Build a simple knowledge-base Q&A system with low-code tools
• Know when to use RAG and how to evaluate its effectiveness

Required tools:

• Large model inference tools (Ollama or online APIs)
• Vector databases (ChromaDB, FAISS, etc.)
• RAG platforms (MaxKB, Dify, and other low-code tools)

Local deployment requirements:

• GPU: RTX 3060 12GB can run a 7B quantized model
• Software stack: Ollama + vector database + RAG frontend
• Difficulty: Medium — requires understanding the collaboration among multiple components

Module 9: AI Content Creation Workflows ​

What to learn:

• AI-assisted writing workflow from topic to finished draft
• AI video workflow from script to finished cut
• AI design workflow from concept to design draft
• Multi-tool collaborative content production
• Human review and quality control of AI outputs
• Copyright compliance and content authenticity

What learners can do after completion:

• Design AI-assisted workflows suited to their own role
• Use AI toolchains to boost content output efficiency
• Establish a quality control mechanism for AI content

Required tools:

• A combination of the tools from the modules above
• Workflow orchestration tools

Local deployment requirements: Depends on which tools the specific workflow uses

III. Development Track: AI Application Development Modules ​

This layer targets learners with programming skills — computer science students, programmers, and technical roles. The goal is to build AI applications from scratch, understand underlying principles, and master the full chain from development to deployment.

Module 10: Large Model Fundamentals and Principles ​

What to learn:

• Transformer architecture and the attention mechanism
• Tokenization and vocabularies
• The training process of large models (pretraining, fine-tuning, alignment)
• Mainstream large model families (GPT, Llama, Qwen, Mistral, DeepSeek, etc.)
• Basic concepts of inference optimization (quantization, KV cache, batching)
• Model evaluation methods and benchmarks

What learners can do after completion:

• Understand how large models work
• Follow the core concepts in technical documentation and papers
• Choose appropriate models based on requirements
• Interpret model evaluation leaderboards and performance data

Required tools:

• Python environment
• Hugging Face Transformers
• Model inference framework

Local deployment requirements:

• GPU: RTX 4070/5070+ recommended (a 7B quantized model can also run on a 3060); for research-grade work, server GPUs such as A100
• Models: at least one 7B-class model for experimentation
• Difficulty: Medium-high — requires Python and deep learning basics

Module 11: Advanced Prompt Engineering and Evaluation ​

What to learn:

• Advanced prompt techniques (chain-of-thought, tree-of-thought, few-shot learning)
• Context engineering and token management
• Design and reuse of prompt templates
• Output format control (JSON, XML, Markdown)
• Systematic evaluation methods for prompts
• Automated prompt optimization

What learners can do after completion:

• Design prompt strategies for complex tasks
• Establish a prompt evaluation and iteration process
• Design reliable output formats for AI applications

Required tools:

• Large model API or local model
• Python (for automated evaluation)

Local deployment requirements: Same as Module 10

Module 12: RAG System Development ​

What to learn:

• Full RAG architecture and components
• Document parsing and processing (PDF, Word, web pages, multiple formats)
• Text chunking strategies (fixed chunking, semantic chunking, recursive chunking)
• Embedding models and vectorization
• Vector database selection and use (Chroma, FAISS, Pinecone, Weaviate)
• Retrieval strategies (keyword retrieval, vector retrieval, hybrid retrieval, reranking)
• RAG evaluation methods (RAGAS, etc.)
• Agentic RAG and advanced RAG architectures

What learners can do after completion:

• Build a RAG Q&A system from scratch
• Optimize retrieval effectiveness for different scenarios
• Evaluate and iterate on the quality of RAG systems
• Design complex RAG application architectures

Required tools:

• Large model inference framework
• Vector database
• RAG frameworks such as LangChain / LlamaIndex
• Document processing libraries

Local deployment requirements:

• GPU: RTX 4070/5070 12GB+; labs can configure higher
• Full software stack: large model + embedding model + vector database + RAG framework
• Difficulty: High — many components; complex debugging

Module 13: AI Agent Development ​

What to learn:

• Basic concepts and core components of Agents (planning, memory, tool calling, action)
• The ReAct framework and the thought-action loop
• Function Calling and tool use
• Multi-agent collaboration and orchestration
• Agent frameworks (LangGraph, AutoGen, CrewAI, etc.)
• Model Context Protocol (MCP)
• Agent evaluation and debugging

What learners can do after completion:

• Design and implement single-agent applications
• Build multi-agent collaboration systems
• Integrate external tools and APIs for Agents
• Debug and optimize Agent performance

Required tools:

• Agent frameworks (LangGraph, etc.)
• Large model inference
• Tool integration SDKs

Local deployment requirements:

• GPU: Same as the RAG module
• Additional dependencies: Agent frameworks, tool integration libraries
• Difficulty: High — abstract concepts; difficult debugging

Module 14: Model Fine-tuning ​

What to learn:

• Why fine-tune? When should you fine-tune?
• Parameter-efficient fine-tuning (PEFT) techniques (LoRA, QLoRA, IA3, etc.)
• Preparing and cleaning fine-tuning data
• The fine-tuning training process and hyperparameter tuning
• Evaluation and testing of fine-tuned models
• Model merging and export
• Multi-turn conversation fine-tuning

What learners can do after completion:

• Use LoRA to fine-tune open-source models for specific domains
• Prepare and clean fine-tuning datasets
• Evaluate fine-tuning effectiveness
• Deploy and use the fine-tuned models

Required tools:

• Fine-tuning frameworks (LLaMA-Factory, PEFT, etc.)
• Training GPUs
• Dataset processing tools

Local deployment requirements:

• GPU: RTX 4090/5090 24GB minimum (for 7B model LoRA fine-tuning); with ample budget, 4090D/5090 or professional compute cards
• The more VRAM the better; 24GB+ recommended; for research-grade work, A100/A6000, etc.
• Training time: several hours to several days
• Difficulty: High — requires deep learning basics; high VRAM requirements

Module 15: Model Quantization and Inference Optimization ​

What to learn:

• Basic principles of model quantization (INT8, INT4, FP8, etc.)
• Comparison of quantization formats (GGUF, AWQ, GPTQ, FP8)
• Using quantization tools
• Assessing the impact of quantization on model quality
• Inference acceleration techniques (KV cache, batching, speculative decoding)
• Inference optimization strategies for different hardware

What learners can do after completion:

• Choose the right quantized version for the hardware
• Quantize a model themselves
• Optimize inference speed and VRAM usage
• Run larger models on limited hardware

Required tools:

• Quantization tools (llama.cpp, AutoGPTQ, etc.)
• Inference frameworks (Ollama, vLLM, SGLang, etc.)
• Model benchmarking tools

Local deployment requirements:

• GPU: RTX 3060+ is sufficient (quantized models are smaller)
• Difficulty: Medium-high — requires understanding hardware and low-level principles

Module 16: AI Application Deployment and Operations ​

What to learn:

• Choosing between local and cloud deployment
• Docker containerized deployment
• API service packaging (FastAPI, etc.)
• Frontend interface development (Gradio, Streamlit, etc.)
• Performance monitoring and logging
• Multi-user concurrency and resource management
• Model updates and version management
• Security and access control

What learners can do after completion:

• Package and deploy AI applications
• Provide API interfaces for other systems to call
• Monitor and maintain AI services
• Manage multi-user environments

Required tools:

• Docker
• API frameworks (FastAPI, Flask, etc.)
• Web UI frameworks (Gradio, Streamlit, etc.)
• Monitoring tools

Local deployment requirements:

• Development environment: a personal computer is sufficient
• Production deployment: a server or cloud host
• Difficulty: Medium-high — requires backend development and ops knowledge

Module 17: AI Security and Compliance Development ​

What to learn:

• Security risks of large models (hallucination, injection attacks, data leakage)
• Prompt injection defense
• Output content filtering and security auditing
• Data privacy protection methods
• Compliance requirements for AI applications
• Open-source model licenses and commercial-use restrictions

What learners can do after completion:

• Identify security risks in AI applications
• Implement basic security protection measures
• Ensure AI applications comply with data privacy regulations
• Choose the right license for open-source models

Required tools:

• Security assessment tools
• Content filtering libraries

Local deployment requirements: No special requirements

How to Use This Catalog ​

1. First, determine the audience and depth ​

Is it AI literacy education for all students? An AI creation course for design majors? Or an AI application development course for computer science majors? Different goals call for entirely different modules.

2. Pick and combine from the module library ​

A few common combination examples:

3. Adjust based on hardware ​

Not enough GPUs? That's fine —

• Literacy modules can be taught entirely with online tools
• Application modules can start with cloud services and gradually transition to local
• The introductory parts of development modules can also use online Colab environments

But we want to emphasize: if conditions allow, introduce local practice as early as possible. With only online tools, students will never truly understand model size, VRAM limits, or the meaning of quantization — these are all very important concepts in AI application development.

4. The practice environment is foundational infrastructure ​

No matter which modules you choose, one thing is common: building and managing the practice environment is the foundational infrastructure of every AI course.

Installing models, configuring dependencies, managing versions, unifying environments — these are not teaching content, but they directly affect teaching quality. Teachers spending large amounts of time helping students troubleshoot environment issues, students giving up because they can't install the tools — these are real wastes in AI education.

The market already has some tools trying to solve this problem — Ollama simplifies running large models locally, ComfyUI turns image generation into a node-based interface, and there are one-stop AI application deployment tools that support rapid deployment of common models in their original, quantized, and accelerated versions. For schools and training institutions, the value of such tools is: hand environment setup off to specialized tools, so teachers spend their time on teaching and students spend their time on learning.

Finally: The Value of Modularity ​

Why make it modular instead of a fixed course? Because AI technology changes too fast — today's hot model may be obsolete next year, today's frontier technology may become basic next year. If the course is "one big block," every update means starting over.

The benefits of modularity:

• Flexible — different audiences pick different modules; combine as needed
• Updatable — when a technology updates, only swap the corresponding module; no need to rewrite the whole course
• Extensible — when a new technology emerges, just add a new module
• Verifiable — each module has clear learning objectives and outputs, making it easy to assess effectiveness

Generative AI course design is essentially not "designing a course" but "building a course system that can iterate continuously." Modules are building blocks; how you stack them and how high you stack them depends on your goals and resources. But no matter how you stack them, one thing is certain — the practice environment is the foundation. Once the foundation is solid, the structure above can stand firm.