Conventional GAN vs Conditional GAN comparison
GANcGANComparisonConditioning
← Generative AIQ34
Compare conventional GANs and Conditional GANs in terms of architecture, input representation, training objective, and control over generated samples. Explain with a suitable application scenario.
What's actually being asked
A structured, side-by-side comparison across four named axes, closed out with one concrete application showing why the difference matters.
Answer
Comparison
| Aspect | Conventional GAN | Conditional GAN (cGAN) |
|---|---|---|
| Generator input | noise z only | noise z + condition y (concatenated or embedded) |
| Discriminator input | sample x only | sample x + the same condition y |
| Architecture | G: z → data space; D: x → score | G: (z,y) → data space; D: (x,y) → score — an extra concatenation/embedding step in both, otherwise the same FC or conv backbone |
| Training objective | min_G max_D 𝔼_{x∼p_data}[log D(x)] + 𝔼_{z∼p_z}[log(1 − D(G(z)))] | min_G max_D 𝔼[log D(x|y)] + 𝔼[log(1 − D(G(z|y)|y))] |
| What D judges | "is x real?" | "is x real for this y?" (joint realism + correctness) |
| What's learned | marginal p(x) | conditional p(x|y) for every y |
| Control at inference | none — sample is random from p_data | choose y, get a sample matching that condition |
| Typical failure modes | mode collapse, vanishing gradients, instability | the same failure modes still apply — conditioning adds control, not extra stability |
Application scenario
Image-to-image translation (e.g. turning a sketch into a photorealistic image, or a daytime photo into its nighttime version) needs a cGAN, not a conventional GAN. Here the condition y is the source image itself (the sketch or day photo), and the output needs to be specifically tied to that input — a conventional GAN sampling from the marginal distribution of "all possible photos" gives no way to request "the photo corresponding to this sketch." The cGAN's D(x|y) explicitly checks that the generated photo is both realistic and consistent with the given sketch, which is exactly the controllability a conventional GAN cannot provide.
Simple shabdo mein: conventional GAN sirf "kuch bhi real-looking" generate karta hai; cGAN "specifically yeh wala, jo di gayi condition se match kare" generate karta hai.
Exam tip
When asked to "compare," write the training objective with y explicit in both D(·|y) terms, not just in prose — graders specifically check that you know where y enters the formula, not just that "a label is added somewhere."