TFG-Flow: Training-free Guidance in Multimodal Generative Flow
This work addresses a gap in training-free guidance for multimodal data, which is incremental but important for scientific applications like drug design.
The paper tackles the problem of generating samples with desired properties from unconditional generative models without additional training, specifically for multimodal data involving both continuous and discrete variables, and demonstrates its effectiveness on molecular design tasks with potential applications in drug design.
Given an unconditional generative model and a predictor for a target property (e.g., a classifier), the goal of training-free guidance is to generate samples with desirable target properties without additional training. As a highly efficient technique for steering generative models toward flexible outcomes, training-free guidance has gained increasing attention in diffusion models. However, existing methods only handle data in continuous spaces, while many scientific applications involve both continuous and discrete data (referred to as multimodality). Another emerging trend is the growing use of the simple and general flow matching framework in building generative foundation models, where guided generation remains under-explored. To address this, we introduce TFG-Flow, a novel training-free guidance method for multimodal generative flow. TFG-Flow addresses the curse-of-dimensionality while maintaining the property of unbiased sampling in guiding discrete variables. We validate TFG-Flow on four molecular design tasks and show that TFG-Flow has great potential in drug design by generating molecules with desired properties.