Course contentsShow
← Machine Learning and Deep Learning
Lesson 1481 of 3,53833. Generative Adversarial NetworksFree lesson

Evaluating GAN Quality

Challenges in evaluation and basic metrics like visual inspection and discriminator accuracy.

Evaluating GAN Quality

What you'll learn: You'll understand why evaluating GANs is uniquely challenging and explore basic approaches to assess the quality of generated samples.

The Evaluation Problem

Unlike supervised learning where you have clear ground truth labels and can compute straightforward accuracy metrics, GANs generate entirely new samples. How do you measure "good" when there's no single correct answer? This is the fundamental challenge: there's no obvious loss function that tells you if your GAN is producing high-quality, diverse images.

Visual Inspection

The most intuitive approach is simply looking at generated samples. Are the images sharp and realistic? Do faces look natural, or are there strange artifacts? While subjective, this "eyeball test" is often your first diagnostic tool.

Analogy: Like evaluating a painter's work—you know good art when you see it, even if you can't quantify exactly why.

Discriminator Accuracy as a Signal

Another basic metric is monitoring your discriminator's accuracy during training. If it classifies real vs. fake perfectly (near 100%), your generator is failing. If it's near random chance (50%), ideally your generator has learned to fool it completely—though this could also mean your discriminator is too weak.

The catch: A 50% discriminator accuracy doesn't guarantee quality. Your generator might have collapsed to producing just one convincing image repeatedly (mode collapse), or both networks might be poorly trained.

Why This Is Hard

These basic metrics can't tell you if your GAN suffers from mode collapse, captures the full diversity of real data, or produces meaningful variations. Visual inspection doesn't scale, and discriminator accuracy is an indirect, noisy signal of generation quality.

Key Takeaway: Evaluating GAN quality is challenging because there's no ground truth for creativity; basic approaches like visual inspection and discriminator accuracy provide initial signals but have significant limitations in capturing true generation quality and diversity.