New paper: learn CFG guidance schedules adversarially with GAN + RL

kwangmoo_yi · x · 2026-08-18

A new arXiv paper, "Adversarial Learning of Classifier-Free Guidance Schedules" (Pokle, Galashov et al., including DeepMind researchers Doucet and De Bortoli), tackles the classic CFG problem in text-to-image diffusion: a static, global guidance scale is suboptimal across timesteps, samples, and conditions. The authors frame learning time-, condition-, and sample-dependent guidance schedules as density ratio estimation: a discriminator estimates the time-dependent log-density ratio between the true and guided marginals, while a lightweight generator network predicts the optimal state-dependent guidance scale. Empirically, the method beats heuristic CFG schedules and prior dynamic-guidance learning approaches on text-to-image benchmarks. The sharer quips: are we going back to adversarial training?

Related event: DeepMind Paper: Adversarial Learning of CFG Guidance Schedules(2 posts)→

Original post →

More from Multimodal

Multimodal channel →