Researchers have introduced Direct Trajectory Optimization (DTO), a novel framework designed to align text-to-image diffusion models with specific reward functions without requiring computationally expensive backpropagation. This training-free approach directly optimizes the sampling trajectory by treating intermediate denoising scores and sampled noise as variables. To address potential distribution issues, DTO incorporates Reward-Optimized Score Guidance, which blends optimized scores with pretrained model predictions to ensure stable sampling and theoretically guaranteed reward improvements. Empirical results demonstrate DTO's effectiveness across multiple diffusion models and benchmarks, outperforming existing inference-time alignment methods while maintaining high efficiency. AI
IMPACT This new framework offers a more efficient method for controlling generative AI outputs, potentially improving user experience and creative applications.
RANK_REASON The cluster contains a research paper detailing a new method for aligning diffusion models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
