ReNO is a novel method for enhancing text-to-image (T2I) models by optimizing initial noise based on human preference reward models. This approach improves the quality and faithfulness of generated images without requiring extensive training. ReNO optimizes the initial noise vector through gradient ascent, using a combination of reward models to avoid reward hacking and ensure better alignment with the input prompt. The method is efficient, achieving significant improvements in performance across various benchmarks, including T2I-CompBench and GenEval, with results that surpass existing open-source models. ReNO requires only 20-50 seconds of computation, making it suitable for practical applications. The method is effective in enhancing the performance of one-step T2I models, with results showing that ReNO-enhanced models outperform existing models in both quantitative and qualitative evaluations. The approach also highlights the importance of noise distribution in T2I models and encourages further research into optimizing this aspect. ReNO provides a practical and efficient solution for improving T2I generation at inference time, offering a compelling trade-off between performance and inference speed.ReNO is a novel method for enhancing text-to-image (T2I) models by optimizing initial noise based on human preference reward models. This approach improves the quality and faithfulness of generated images without requiring extensive training. ReNO optimizes the initial noise vector through gradient ascent, using a combination of reward models to avoid reward hacking and ensure better alignment with the input prompt. The method is efficient, achieving significant improvements in performance across various benchmarks, including T2I-CompBench and GenEval, with results that surpass existing open-source models. ReNO requires only 20-50 seconds of computation, making it suitable for practical applications. The method is effective in enhancing the performance of one-step T2I models, with results showing that ReNO-enhanced models outperform existing models in both quantitative and qualitative evaluations. The approach also highlights the importance of noise distribution in T2I models and encourages further research into optimizing this aspect. ReNO provides a practical and efficient solution for improving T2I generation at inference time, offering a compelling trade-off between performance and inference speed.