Training-Free Guidance Beyond Differentiability: Scalable Path Steering with Tree Search in Diffusion and Flow Models

2025-02-17

AI Summary

Training-free guidance in diffusion and flow models fails when objectives are non-differentiable or data is discrete, undermining conventional gradient-based guidance. Method: This paper introduces TreeG—a novel training-free guidance framework that unifies guidance across continuous and discrete domains by integrating tree search. TreeG generates candidate paths, evaluates them via non-differentiable objectives, and selects the optimal path without relying on gradients. It supports parallel candidate exploration and path distillation, and reveals a scaling law between inference compute and performance. Contributions/Results: TreeG outperforms all existing state-of-the-art training-free methods on symbolic music generation, small-molecule design, and enhancer DNA sequence generation. Empirical evaluation demonstrates its high efficiency and strong scalability in inference, validating both effectiveness and computational tractability across diverse generative tasks.

Technology Category

• Search and Optimization: Sampling/Simulation-based Search
• Machine Learning: Deep Generative Models & Autoencoders
• Computer Vision: Diffusion Models for Vision

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Training-free guidance enables controlled generation in diffusion and flow models, but most existing methods assume differentiable objectives and rely on gradients. This work focuses on training-free guidance addressing challenges from non-differentiable objectives and discrete data distributions. We propose an algorithmic framework TreeG: Tree Search-Based Path Steering Guidance, applicable to both continuous and discrete settings in diffusion and flow models. TreeG offers a unified perspective on training-free guidance: proposing candidates for the next step, evaluating candidates, and selecting the best to move forward, enhanced by a tree search mechanism over active paths or parallelizing exploration. We comprehensively investigate the design space of TreeG over the candidate proposal module and the evaluation function, instantiating TreeG into three novel algorithms. Our experiments show that TreeG consistently outperforms the top guidance baselines in symbolic music generation, small molecule generation, and enhancer DNA design, all of which involve non-differentiable challenges. Additionally, we identify an inference-time scaling law showing TreeG's scalability in inference-time computation.

Problem

• Training-free guidance for non-differentiable objectives
• Scalable path steering in diffusion models
• Tree search-based approach for discrete data

Innovation

• Tree search-based path steering
• Handles non-differentiable objectives
• Unified training-free guidance framework