In a conditional GAN, why does the discriminator receive the class label too?

## What "conditional" means here An unconditional generator learns to sample from the data distribution `p(x)` — give it noise, get a plausible image. A conditional generator learns the family of distributions `p(x|y)`, where `y` is a side input: a class label from a taxonomy, an attribute vector, a source image, a text embedding, or a dense time-aligned signal. The contract is that the caller chooses `y` and gets a sample that is both realistic *and* consistent with `y`. The original conditional formulation is deliberately minimal: feed `y` to both players. The generator computes `G(z, y)`; the discriminator computes `D(x, y)`. Nothing else about the adversarial setup changes. ## Why one-sided conditioning fails Suppose you feed `y` only to the generator and leave the discriminator judging images alone. The discriminator's job is then to separate real images from generated images, full stop. Its verdict on a sample does not depend on the label that produced it, so the generator's training signal does not depend on that label either. Given that signal, the generator's best achievable behaviour is to produce samples indistinguishable from the training set *overall*. It is not rewarded for lining up class 7 with images of class 7, and honouring the label costs it capacity. So the extra input becomes a second, structureless noise channel. You end up with an unconditional model wearing a conditional interface: the sample is realistic, the label is decoration, and nothing in the loss curves tells you. A classifier bolted on after the fact does not fix this either — it changes the objective, but it does not change the fact that the discriminator, the component that defines what "real" means, is blind to `y`. ## What the conditional discriminator actually enforces With `D(x, y)`, the positive examples are *real pairs* drawn from the joint distribution: an image together with the label it genuinely carries. Everything else is negative — generated pairs, and (in some formulations, explicitly) real images paired with a shuffled label. The discriminator therefore has to learn what makes an image and a label go together, and the generator has to satisfy that joint judgement. The equilibrium it is being pushed toward is the joint `p(x, y)`, which is exactly `p(x|y)` for the label distribution you sample from. That is the whole idea: label adherence is enforced by the discriminator, not requested by the architecture. ## How the label physically enters each player For the generator, the crudest route is to embed the label into a vector and concatenate it with the noise vector at the input. This works but is weak — a single injection at the bottom of a deep stack is easy for later layers to wash out. Stronger routes inject the conditioning repeatedly: predict the per-channel scale and shift of a normalisation layer from the label (conditional normalisation), so every block is reminded of `y`. For a dense conditioning signal such as a spectrogram frame sequence or a source image, the conditioning is spatially or temporally aligned with the output and is injected as extra channels at matching resolutions. For the discriminator, the naive route is to tile the label embedding into a constant feature map and concatenate it to the input or to an intermediate activation. This is workable for a handful of classes and gets weak as the taxonomy grows, which is why large class-conditional models use structured alternatives that make the label interact multiplicatively with the image features rather than sitting beside them. ## Detecting a dead condition Three cheap checks, none of which requires new machinery: 1. **Label sweep with fixed noise.** Freeze `z`, iterate `y` over the taxonomy. In a working model the identity of the content changes while style and pose stay related. In a broken one, you get the same picture every time. 2. **Noise sweep with fixed label.** The complement: vary `z` with `y` pinned. Outputs should stay on-label but differ from one another; if they are near-duplicates, the label dominates and diversity has gone. 3. **Independent judge.** Train a classifier on *real* data only, then classify a large batch of samples and compare the predicted label against the requested one. A confusion matrix that is far from diagonal is direct evidence that the conditioning is not being honoured. ## Forcing the label to matter If a sweep shows the label is ignored, the fixes are ordered: confirm the discriminator actually receives `y` at all; strengthen the injection path in the generator so conditioning enters at multiple depths rather than once; and add explicitly mismatched real pairs to the discriminator's negative set so "wrong label" is a first-class way to be fake. Only after those should you reach for auxiliary losses — they trade label adherence against sample diversity, and paying that price to patch a plumbing bug is the wrong trade.