P5-15.6 How Does a VAE Turn an Image into a Latent Representation?¶
Section ID:
P5-15.6Version:v2026.08.28
P5-15.4 presented diffusion as an algorithm that repeatedly restores noisy states, and P5-15.5 presented U-Net and DiT as choices for predicting that restoration direction. Latent diffusion performs repeated computation in a latent representation rather than directly in pixels. A VAE (variational autoencoder) family connects an image and that latent representation.
The question in this section is how a VAE differs from an ordinary autoencoder, why it makes a latent space usable for generation, and which calculation it performs in latent diffusion.
An Autoencoder Learns a Representation That Can Be Reconstructed¶
An ordinary autoencoder maps an input image x to an often more compact representation z through an encoder, then uses a decoder to reconstruct a similar image x_hat. Its central signal is the difference between input and reconstruction. A latent representation does not always mean fewer dimensions; here we first read it as a representation space suitable for repeated generation computation.
| Component | What it does | First question to ask |
|---|---|---|
| encoder | maps image x to latent z | what information remains in a smaller representation? |
| latent representation | intermediate coordinates left by the encoder | how might similar inputs be arranged? |
| decoder | reconstructs x_hat from z | how much of the original scene can the retained information restore? |
| reconstruction loss | measures the difference between x and x_hat | does learning preserve important input information? |
This is enough to learn reconstructable representations. But if the latent coordinates made by the encoder are scattered with large gaps, a randomly chosen coordinate may not decode into a natural image. Reconstructing well and generating from a new coordinate are different requirements.
A VAE Produces a Distribution Rather Than One Fixed Coordinate¶
Instead of directly returning one latent coordinate for each image, a VAE encoder predicts a mean mu and spread sigma for a distribution associated with the image. A sample z from that distribution is sent to the decoder.
flowchart TD
A[Image x] --> B[VAE encoder]
B --> C[mu and sigma]
C --> D[latent sample z]
D --> E[Diffusion: noise prediction and repeated restoration]
E --> F[final latent]
F --> G[VAE decoder]
G --> H[generated image]
| VAE value | Meaning | Common confusion |
|---|---|---|
mu | center of the latent distribution for this input | not a finished image |
sigma | spread around that center | not a quality score or predicted noise |
z | latent sample made from mu, sigma, and randomness | not the same as a time-step diffusion noise state |
x_hat | image reconstructed by the decoder from z | not always a diffusion model's final generated result |
During training, z is commonly constructed with random epsilon sampled from a standard normal distribution.
This is the reparameterization idea: sampling remains random, while the loss can still affect the encoder that produces mu and sigma. This epsilon creates a VAE latent sample; it is not the diffusion noise added to an image state at a time step in P5-15.4.
Reconstruction and KL Loss Protect Different Requirements¶
A VAE training objective can be read as reconstruction loss plus a KL-divergence term.
The useful point is to distinguish the failure each term prevents.
| Loss term | Problem it reduces | Risk when its influence is poorly balanced |
|---|---|---|
| reconstruction loss | decoder loses important content or structure | weak reconstruction can blur outputs; overemphasis can leave irregular scattered coordinates |
| KL divergence | each input distribution drifts too far from a common normal reference | too much pressure can remove information needed for reconstruction |
The KL term does not force every image into one point. It links input-specific distributions to a reference distribution from which generation can sample. A VAE is therefore more than a compression tool: it aims for a latent space whose nearby coordinates remain usable.
VAE and Diffusion Are Different Stages in Latent Diffusion¶
Latent diffusion runs its noise prediction and repeated restoration in the latent representation made by a VAE. The VAE moves between image and representation; the diffusion denoiser predicts a noise direction inside the latent state.
| Stage | Main model | Input to output | Question it answers |
|---|---|---|---|
| latent encoding | VAE encoder | image x -> latent z | how do we move an image to a workable representation? |
| latent generation | diffusion denoiser and scheduler | noisy latent -> restored latent | what should be removed at this step and how do we move next? |
| image decoding | VAE decoder | final latent -> image | how do we turn the latent result into a visible image? |
A VAE is neither a diffusion scheduler nor a U-Net or DiT denoiser, and not every diffusion model uses one. Pixel-space diffusion is also possible. Latent diffusion is one design that changes the space in which repeated computation happens.
Check the Boundary with a Small Comparison¶
| Statement | Component | Why |
|---|---|---|
turn an image into a latent representation | VAE encoder | representation conversion before repeated diffusion |
predict noise from the current noisy latent | U-Net or DiT | diffusion restoration-direction prediction |
calculate the next latent state from the prediction | scheduler | generation-path rule, not learned weights |
turn the final latent into pixels | VAE decoder | makes the result visible to people |
If these statements are distinguishable, we can avoid treating different calculations as one claim such as the VAE makes the image or diffusion trains the VAE.
Checklist¶
- I can explain how an autoencoder uses an encoder, latent representation, decoder, and reconstruction loss.
- I can explain why a VAE encoder makes a latent distribution through
muandsigma. - I can distinguish the jobs of reconstruction loss and KL divergence.
- I can keep VAE sampling
epsilonseparate from time-step diffusion noise. - I can separate the VAE encoder, diffusion denoiser and scheduler, and VAE decoder in latent diffusion.
Sources and Further Reading¶
- Diederik P. Kingma, Max Welling, Auto-Encoding Variational Bayes, ICLR, 2014, accessed 2026-08-28.
- Robin Rombach et al., High-Resolution Image Synthesis with Latent Diffusion Models, CVPR, 2022, accessed 2026-08-28.