ReNO
ReNO is an inference-time enhancement method for one-step text-to-image models that optimizes initial noise using reward signals from human preference models. Presented at NeurIPS 2024, this approach addresses the difficulty of capturing intricate details in complex prompts that often plagues diffusion models. Unlike traditional fine-tuning which can suffer from reward hacking and poor generalization, ReNO performs gradient ascent on the noise input for approximately 50 iterations, typically taking 20 to 50 seconds of computation. This process allows one-step models like SD-Turbo, SDXL-Turbo, PixArt, and Hyper-SD to surpass the quality of current open-source competitors such as SDXL and PixArt-alpha within the same computational budget. Benchmarks including T2I-CompBench and GenEval demonstrate its superior performance, while user studies indicate a strong preference over SDXL and parity with the proprietary 8B parameter Stable Diffusion 3. The software is designed for researchers and developers seeking to bo