Someone actually baked FID directly into the training process and re-ran ImageNet.

Town Square / 小镇广场
1

There’s this rumor going around that’s pretty intriguing—saying OpenAI’s involved too, and they’re reviving ImageNet by making FID the actual training objective. FID’s always been the old-school metric for evaluating image generation quality, usually you just calculate it after training to see how you did.

But now they’re straight-up plugging it into the training loop as the target to optimize for, which is pretty counterintuitive. I haven’t found the original paper on how exactly they pulled it off, but the direction feels like they want the model to align with the evaluation standard during training itself.

Also, ImageNet getting dusted off and thrown back into the ring is kinda interesting—old benchmark getting a second wind. Anyone here managed to dig up the paper? Drop a link, please.

2

Yeah, baking evaluation metrics directly into the loss function definitely makes it way too easy to overfit to the metric itself.

3

FID’s always been controversial anyway, directly optimizing for it isn’t necessarily a good thing.

4

Oh great, ImageNet’s getting dragged out for another round of public execution.

5

Can someone drop the original post? The title doesn’t really give away the good stuff.

6

OpenAI’s name gets slapped on a lot of stuff, but that doesn’t mean they’re the ones calling the shots.