With the rise of Stable Diffusion and Midjourney, it seems like GANs are becoming obsolete for image generation. Are there still use cases where a GAN is preferred over a Diffusion model, especially when considering inference speed and hardware requirements for production?
3 answers
Diffusion models definitely win on image quality and "creativity" because they don't suffer from mode collapse as easily as GANs. However, GANs are still the kings of real-time inference. A GAN can generate an image in a single forward pass, whereas a Diffusion model has to iteratively denoise the image over 20-50 steps. If you are building an app that needs to generate images instantly (like a real-time video filter), a GAN or a specialized "Latent Consistency Model" (which speeds up diffusion) is your only realistic choice for 2025.
Does the training cost factor into your decision? I've heard that training a Diffusion model from scratch requires significantly more data and VRAM.
GANs are still widely used in "Image-to-Image" translation tasks like style transfer or super-resolution where the input structure is already defined.
Agreed. For medical imaging or satellite enhancements, the speed and structural consistency of GANs like Pix2Pix are still very relevant.
You're right, Patrick. Training a foundation Diffusion model is a massive undertaking. However, most people just use ControlNet or LoRA to adapt existing models, which is much more manageable. For most startups, the "pre-trained" ecosystem of Diffusion is far more valuable than starting from zero with a GAN.