New research demonstrates generative adversarial networks can scale effectively when paired with transformer architectures

@jm_alexia did they just take https://arxiv.org/abs/2107.04589 and scaled it?

@jm_alexia Cooking something even better right now, just wait. :)

@jm_alexia Wow, these are some brave researchers :)

@jm_alexia love me a good GAN because I enjoy imagining the models fight each other

I think too, having it as a hybrid objective similar to mixing it in with VAE reconstruction losses or even tricks in the realm of GRPO feel helpful here

@jm_alexia Love seeing researchers revisit ideas instead of chasing hype only

@jm_alexia 🍿🧐 interesting transformer GANs in latent space feels like GANs might have a path again :3

@jm_alexia woohoo. Time for model collapse

@jm_alexia Type of paper i like. When you show some older method is still relevant. 👌 Though i love diffusion but it seems GANs are somehow unkillable.

@jm_alexia interesting. Sent you a DM about your research workflow. Please check it when you get a moment.

@jm_alexia GANs were the simplest form of training method. I loved creating them.

@jm_alexia why not cite mUP?

@jm_alexia GAN GAN GAN

@arataeb @jm_alexia From what skimmed, they modified relativistic gan loss and turned MSG into MNG. A couple out of many more for sure.

@jm_alexia i will forever love gans

Minimizing something like MMD against the real reference distribution (either against single sample or generated distribution) feels like a safer route somewhat for the goal of real/fake distinction and blending in at a distribution level

@jm_alexia Guess they were GAN but not forgotten

@SkyLi0n @jm_alexia 👀

@SkyLi0n @jm_alexia Hype! 🙌

@jm_alexia