What's this about?
Fréchet distance has become an effective training objective for improving image and video generators after their main training run. The problem this paper (23 upvotes) addresses is "Fréchet hacking": optimizing the metric directly can make the score improve while the actual generated images get worse. AdvFD adds an adversarial component so the target keeps moving and can't be gamed.
What exactly is Fréchet distance here?
It's a way of scoring a generator on the *whole distribution* of what it produces rather than on individual samples: take thousands of real images and thousands of generated ones, summarize each set statistically, and measure how far apart those summaries are. It's the basis of FID, the standard image-generation metric. The distinction that matters is distribution-level versus sample-level: a generator that produces one perfect image but no variety scores badly, which is exactly what you want.
Why is this worth noticing?
- It's a clean instance of Goodhart's law in machine learning — a measure that stops being a good measure once it becomes a target — and a concrete defense against it.
- Post-training refinement of generators is where much of the current quality improvement in image and video models comes from, so the objectives used there directly shape output quality.