Diffusion has replaced GANs in image generation, and we admit this call can barely lose — resolves 2027-07-14

AI image generation has two main technical routes, GANs (generative adversarial networks, two models trained against each other) and diffusion models, which build an image step by step out of noise, and since 2022 diffusion has replaced GANs as the mainstream.

Which route leads decides where research and computing investment in image generation concentrates.

SecondSource holds that diffusion’s win is settled, and the opposing case needs GANs to return to the mainstream in general-purpose generation, which we rate as very unlikely.

SecondSource’s reading: in visual generation, diffusion replacing GANs has been settled since 2022. If, at the July 2027 check, GANs remain confined to specialist niches such as super-resolution (sharpening low-resolution images), the call is right.

In fairness: when this card was written we put the chance of a reversal below 5%, which makes it a call that can barely lose. It stays public for readers to check, but it is excluded from our hit rate and from weighted scores.

Ho, Jain, Abbeel — UC Berkeley — arXiv
“On the unconditional CIFAR10 dataset, we obtain an…”

The structure is settled. Checkpoint: in July 2027, GANs still confined to specialist niches (super-resolution and the like) counts as a hit.

What would overturn it

GAN-family models return to the mainstream in general-purpose generation (extremely unlikely; listed as a formal reversal only).