This course is not just about "how to implement" a Transformer, but about dissecting why this architecture was created, what role each module plays, and how the entire model works from the designer's perspective. We deeply analyze the internal computation principles of Self-Attention and Multi-Head Attention, and directly verify through formulas, papers, and implementation code what limitations Positional Encoding, Feed-Forward Networks, and Encoder·Decoder structures were introduced to solve. Starting from Attention, we assemble the entire Transformer structure ourselves, and actually perform training to experience firsthand how the model operates. This course is the most structured and practical roadmap for "anyone who wants to completely understand Transformers."

We understand why Attention was needed and how it works by 'implementing it directly with code'. This lecture starts from the structural limitations of RNN and Seq2Seq models, experimentally verifies the information bottleneck problem and long-term dependency issues created by fixed context vectors, and naturally explains how Attention emerged to solve these limitations. Rather than simply introducing concepts, we directly confirm RNN's structural limitations and Seq2Seq's information bottleneck problems through experiments, and implement **Bahdanau Attention (additive attention)** and **Luong Attention (dot-product attention)** one by one to clearly understand their differences. Each attention mechanism forms Query–Key–Value relationships in what way, has what mathematical and intuitive differences in the weight calculation process, and why it inevitably led to later models naturally connects to their characteristics and evolutionary flow. We learn how Attention views sentences and words, and how each word receives importance weighting to integrate information in a form where formula → intuition → code → experiment are connected as one. This lecture is a process of building 'foundational strength' to properly understand Transformers, helping you deeply understand why the concept of Attention was revolutionary, and why all subsequent state-of-the-art NLP models (Transformer, BERT, GPT, etc.) adopt Attention as a core component. This lecture is optimized for learners who want to embody the flow from RNN → Seq2Seq → Attention not through concepts but through code and experiments.

We implement the latest Transformer-based PixArt and lightweight adaptation SANA step by step from theory to code. Building on DDPM·DDIM·LDM·DiT covered in Parts I·II, we complete hands-on practice including text encoder integration, samplers (DDIM/ODE), v-prediction/CFG tuning, and small-scale data style fine-tuning.

This course is a hands-on masterclass that completely dissects the core technological evolution of generative AI, from LDM (Latent Diffusion Model) to DiT (Diffusion Transformer). We directly analyze the latent space-based learning principles of LDM, the structure of Stable Diffusion, and the implementation methods of the latest Diffusion Transformer through papers and code. Students will systematically learn the latest trends and structural evolution of generative models by directly implementing LDM, CFG (Classifier-Free Guidance), and DiT models using PyTorch.

This course is a hands-on masterclass that completely conquers the evolution of Diffusion Models through papers and code implementation. You'll learn the core models of generative AI, including DDPM (Denoising Diffusion Probabilistic Model) and DDIM, by studying the paper principles and implementing them directly. We analyze step-by-step the background of each model's emergence, mathematical formulations, network architectures (U-Net, VAE, Transformer), training processes (Noise Schedule, Denoising Step), and the ideas that led to performance improvements. Students will directly code all models using PyTorch, gaining not just paper comprehension but 'practical skills to reproduce and apply' them in real-world scenarios. Additionally, by comparing the differences between models and their developmental flow, you'll clearly understand how they expand and evolve. This course integrates theory, code, and practice into one comprehensive journey, providing researchers, developers, and creators alike with a systematic way to master the evolution of generative models. Beyond simply 'reading' papers, start your experience of 'understanding and recreating' through direct implementation now.