diff --git a/src/year1/deep-learning/img/gan.png b/src/year1/deep-learning/img/gan.png
index 50a32e7..7bfc80a 100644
Binary files a/src/year1/deep-learning/img/gan.png and b/src/year1/deep-learning/img/gan.png differ
diff --git a/src/year1/deep-learning/img/gan_manifold.png b/src/year1/deep-learning/img/gan_manifold.png
new file mode 100644
index 0000000..b52277b
Binary files /dev/null and b/src/year1/deep-learning/img/gan_manifold.png differ
diff --git a/src/year1/deep-learning/img/latent_mapping.png b/src/year1/deep-learning/img/latent_mapping.png
new file mode 100644
index 0000000..a4be46a
Binary files /dev/null and b/src/year1/deep-learning/img/latent_mapping.png differ
diff --git a/src/year1/deep-learning/sections/_generative_models.tex b/src/year1/deep-learning/sections/_generative_models.tex
index 1dae286..5c699c8 100644
--- a/src/year1/deep-learning/sections/_generative_models.tex
+++ b/src/year1/deep-learning/sections/_generative_models.tex
@@ -49,8 +49,10 @@ Generative models are categorized into two families:
 
 
 \section{Variational autoencoder (VAE)}
+\marginnote{Variational autoencoder (VAE)}
+
+Approach belonging to the family of compressive models.
 
-Model belonging to the family of compressive models.
 
 \subsection{Training}
 
@@ -119,16 +121,40 @@ The decoder generates new data by simply taking as input a latent variable $z$ s
 
 
 
-\section{Generative adversarial network (GAN)}
+\subsection{Problems}
 
-Model belonging to the family of compressive models.
-
-During training, the generator is paired with a discriminator that learns to distinguish between real data and generated data.
-
-The loss function aims to:
 \begin{itemize}
-    \item Instruct the discriminator to spot the generator.
-    \item Instruct the generator to fool the discriminator.
+    \item Balancing the two losses is difficult.
+    \item It is subject to the posterior collapse problem, where the model learns to ignore a subset of latent variables.
+    \item There might be a mismatch between the prior distribution and the learned latent distribution.
+    \item Generated images are blurry.
+\end{itemize}
+
+
+
+\section{Generative adversarial network (GAN)}
+\marginnote{Generative adversarial network (GAN)}
+
+Approach belonging to the family of compressive models.
+
+\subsection{Training}
+During training, the generator $G$ is paired with a discriminator $D$ that learns to distinguish between real and generated data.
+
+The loss function is the following:
+\[
+    \mathcal{V}(D, G) = \mathbb{E}_{x \sim p_\text{data}(x)}[\log D(x)] + \mathbb{E}_{z \sim p_z(z)}[\log\big( 1 - D(G(z)) \big)]
+\]
+where:
+\begin{itemize}
+    \item $\mathbb{E}_{x \sim p_\text{data}(x)}[\log D(x)]$ is the negative cross-entropy of the discriminator w.r.t. the true data distribution $p_\text{data}$
+        (i.e. how well the discriminator recognizes real data).
+    \item $\mathbb{E}_{z \sim p_z(z)}[\log\big( 1 - D(G(z)) \big)]$ is the negative cross-entropy of the discriminator w.r.t. the generator
+        (i.e. how well the discriminator is able to detect the generator).
+\end{itemize}
+In other words, the loss aims to:
+\begin{itemize}
+    \item Instruct the discriminator to spot the generator ($\max_D \mathcal{V}(D, G)$).
+    \item Instruct the generator to fool the discriminator ($\min_G \mathcal{V}(D, G)$).
 \end{itemize}
 
 \begin{figure}[H]
@@ -136,11 +162,32 @@ The loss function aims to:
     \includegraphics[width=0.5\linewidth]{./img/gan.png}
 \end{figure}
 
+For more stability, training is done alternately by training the discriminator with the generator frozen and vice versa.
+
+\begin{remark}
+    GANs have the property of pushing the reconstruction towards the natural image manifold.
+    \begin{figure}[H]
+        \centering
+        \includegraphics[width=0.4\linewidth]{./img/gan_manifold.png}
+        \caption{Comparison of GAN and MSE generated images. MSE is obtained as the pixel-wise average of the natural images.}
+    \end{figure}
+\end{remark}
+
+
+\subsection{Problems}
+
+\begin{itemize}
+    \item A generator able to fool the discriminator does not necessarily mean that the generated images are good.
+    \item There are problems related to counting, perspective and global structure.
+    \item The generator tends to specialize on fixed samples (mode collapse).
+\end{itemize}
+
 
 
 \section{Normalizing flows}
+\marginnote{Normalizing flows}
 
-Model belonging to the family of dimension-preserving models.
+Approach belonging to the family of dimension-preserving models.
 
 The generator is split into a chain of invertible transformations.
 During training, the log-likelihood is maximized.
@@ -156,14 +203,93 @@ During training, the log-likelihood is maximized.
 
 
 
-\section{Diffusion models}
+\section{Diffusion model}
+\marginnote{Diffusion model}
 
-Model belonging to the family of dimension-preserving models.
+Approach belonging to the family of dimension-preserving models.
 
-The generator is split into a chain of denoising steps that attempt to remove a Gaussian noise with varying $\sigma$.
-It is assumed that the latent space is a noisy version of the image.
+
+\subsection{Training (forward diffusion process)}
+
+Given an image $x_0$ and a signal ratio $\alpha_t$ (that indicates how much original data is in the noisy image),
+the generator $G$ is considered as a denoiser and a training step $t$ does the following:
+\begin{enumerate}
+    \item Normalize $x_0$.
+    \item Generate a Gaussian noise $\varepsilon \sim \mathcal{N}(0, 1)$.
+    \item Generate a noisy version $x_t$ of $x_0$ by injecting the noise as follows:
+        \[ x_t = \sqrt{\alpha_t} \cdot x_0 + \sqrt{1-\alpha_t} \cdot \varepsilon \]
+    \item Make the network predict the noise $G(x_t, \alpha_t)$ and train it to minimize the prediction error:
+        \[ \Vert \varepsilon - G(x_t, \alpha_t) \Vert \]
+\end{enumerate}
+
+\begin{remark}
+    The values of $\alpha_t$ are fixed by a scheduler.
+\end{remark}
+
+
+\subsection{Inference (reverse diffusion process)}
+
+The generation process is split into a finite chain of $T$ denoising steps that attempt to remove a Gaussian noise with varying $\sigma$.
+(i.e. it is assumed that the latent space is a noisy version of the image).
+
+Given a generator $G$ and a fixed signal ratio scheduling $\alpha_T > \dots > \alpha_1$, an image is sampled as follows:
+\begin{enumerate}
+    \item Start from some random noise $x_T \sim \mathcal{N}(0, 1)$.
+    \item For $t$ in $T, \dots, 1$:
+    \begin{enumerate}
+        \item Estimate the noise using the generator $G(x_t, \alpha_t)$.
+        \item Compute the denoised image $\hat{x}_0$:
+            \[ \hat{x}_0 = \frac{x_t - \sqrt{1-\alpha_t} \cdot G(x_t, \alpha_t)}{\sqrt{\alpha_t}} \]
+        \item Compute a new noisy image for the next iteration by re-injecting some noise with signal ratio $\alpha_{t-1}$ (i.e. inject an amount of noise that is less than this iteration):
+            \[ x_{t-1} = \sqrt{\alpha_{t-1}} \cdot \hat{x}_0 + \sqrt{1 - \alpha_{t-1}} \cdot \varepsilon \]
+    \end{enumerate}
+\end{enumerate}
 
 \begin{figure}[H]
     \centering
     \includegraphics[width=0.65\linewidth]{./img/diffusion_model.png}
-\end{figure}
\ No newline at end of file
+\end{figure}
+
+\begin{remark}
+    A conditional U-net for denoising works well as the generator.
+\end{remark}
+
+
+
+\section{Latent space exploration}
+
+\begin{description}
+    \item[Representation learning] \marginnote{Representation learning}
+        Learning a latent space in such a way that particular changes reflect a desired alteration of the visible space.
+
+        \begin{remark}
+            Real-world data depend on a relatively small set of latent features.
+        \end{remark}
+
+    \item[Disentanglement] \marginnote{Disentanglement}
+        The latent space learned by a model is usually entangled (i.e. a change in one attribute might affect the others).
+
+        Through linear maps, it is possible to pass from one latent space to another.
+        This can be done by finding a small set of points common to the starting and destination spaces (support set)
+        and defining a map based on those points.
+
+        \begin{remark}
+            The latent space seems to be independent of:
+            \begin{itemize}
+                \item The training process.
+                \item The training architecture.
+                \item The learning objective (i.e. GAN and VAE might have the same latent space).
+            \end{itemize}
+        \end{remark}
+
+        \begin{figure}[H]
+            \centering
+            \includegraphics[width=0.3\linewidth]{./img/latent_mapping.png}
+            \caption{
+                \parbox[t]{0.72\linewidth}{
+                    Example of mapping from a latent space $Z_1$ to a space $Z_2$ through $M$.
+                    The two spaces are evaluated on the visible space $V$.
+                }
+            }
+        \end{figure}
+\end{description}
\ No newline at end of file