🎨 Google shows a unified way to control diffusion generators
A Google team (Chih-Wei Hsu, Moonkyung Ryu, and co-authors) published the Diffusion Controller (DiffCon) method on September 29: the reverse sampling of a diffusion model is described as stochastic control in the LS-MDP formalism, and a side network at denoising steps steers the model with a frozen backbone — even a closed one via gray-box access.
🌍 The method unifies generation control techniques (classifier-free guidance, LoRA, reward regression) into a single theory — a template for controllability and safety layers for closed models without access to weights.
👤 Anyone generating images gets an alternative to LoRA: an add-on that learns while touching fewer layers. On Stable Diffusion v1.4 (HPS-v2), gray-box DiffCon outperformed LoRA, and white-box achieved a 90% win rate against the base.
Source 1: https://research.google/blog/how-diffusion-controller-unifies-and-simplifies-ai-image-generation/ Source 2: https://arxiv.org/abs/2603.06981