A^2-TFG: Analytical and Adaptive Training-Free Diffusion Guidance

Chenxin Li, Qinpeng Cui, Yixuan Liu, Dong Zhou, Zhongdao Wang

Keywords: Training-Free Diffusion Guidance, Analytical Solution, Adaptive Guidance

Abstract: Training-free diffusion guidance is a strategy that uses an unconditional diffusion model and an off-the-shelf property predictor to generate samples with desired characteristics without further training. Typical predictors here are any differentiable functions (\eg classifiers) that can be used to evaluate the quality of the generated samples. Existing works design the weights of the guidance term using heuristic rules, resulting in fixed guidance weights for different samples. In this paper, we propose Analytical and Adaptive Training-free Guidance (T^2-AFG), which improves upon prior approaches in two aspects: \textbf{(1) Analytical}: We formulate an optimization objective that provides a closed-form solution for the guidance term weights;\textbf{ (2) Adaptive}: This closed-form solution varies with different inputs rather than being fixed. Compared to heuristic rules or grid search, these improvements lead to generally better performance, and the closed-form solution also reduces computational costs. We extensively validate the effectiveness of T^2-AFG across six tasks combining three models such as Cat-DDPM, Stable Diffusion, and Audio-Diffusion with various task-specific targets, achieving superior performance over the vanilla TFG across most metrics.

Primary Area: generative models

Submission Number: 448