GAN Training Loop

Generator vs Discriminator adversarial loop with labeled losses and gradient flow.

Prompt

A Generative Adversarial Network training loop diagram.

Left — Generator G:
- Takes random noise z ~ N(0, I) as input.
- Outputs a fake sample G(z) drawn as a small image thumbnail.

Right — Discriminator D:
- Takes either a real sample x_real (from the dataset block at the top-right) or G(z).
- Outputs a probability D(x) in [0, 1].

Center — Loss block:
- Adversarial loss for D: L_D = -E[log D(x_real)] - E[log (1 - D(G(z)))].
- Adversarial loss for G: L_G = -E[log D(G(z))] (non-saturating form).

Show two gradient arrows:
- Solid red arrow: D's gradient updates D's weights.
- Solid blue arrow: G's gradient flows back through D into G.

Style: clean publication-style schematic, white background, navy / red / blue palette, sans-serif labels.

Use in Generator

When to use

For generative-modeling papers using a GAN-style training objective.

Variations

Conditional GAN

Same loop but add a class label y as additional conditioning input to both G and D. Show y as a separate input arrow into both networks. Update losses to D(x | y) and G(z | y).

Tips

Color-code G and D gradient arrows differently — readers track gradient flow visually.
Show the data block (real samples) outside the loop so it doesn't crowd the center.
List both losses explicitly. Diagrams without losses miss the training story.

FAQ

Can I show mode collapse in this figure?

Add a small inset chart showing G's sample diversity over training epochs collapsing to a single mode at some step.