TENCENT(HK:00700) 量子位·2025-09-15 03:59Core Viewpoint - The article discusses advancements in AI image generation, specifically focusing on the introduction of two key methods, Direct-Align and Semantic Relative Preference Optimization (SRPO), which significantly enhance the quality and aesthetic appeal of generated images [5][14]. Group 1: Current Challenges in Diffusion Models - Existing diffusion models face two main issues: limited optimization steps leading to "reward hacking," and the need for offline adjustments to the reward model for achieving good aesthetic results [4][8]. - The optimization process is constrained to the last few steps of the diffusion process due to high gradient computation costs [8]. Group 2: Direct-Align Method - Direct-Align method allows for the recovery of original images from any time step by pre-injecting noise, thus avoiding the limitations of optimizing only in later steps [5][10]. - This method enables the model to recover clear images from high noise states, addressing the gradient explosion problem during early time step backpropagation [11]. - Experiments show that even at just 5% denoising progress, Direct-Align can recover a rough structure of the image [11][19]. Group 3: Semantic Relative Preference Optimization (SRPO) - SRPO redefines rewards as text-conditioned signals, allowing for online adjustments without additional data by using positive and negative prompt words [14][16]. - The method enhances the model's ability to generate images with improved realism and aesthetic quality, achieving approximately 3.7 times and 3.1 times improvements, respectively [16]. - SRPO allows for flexible style adjustments, such as brightness and cartoon style conversion, based on the frequency of control words in the training set [16]. Group 4: Experimental Results - Comprehensive experiments on the FLUX.1-dev model demonstrate that SRPO outperforms other methods like ReFL, DRaFT, and DanceGRPO across multiple evaluation metrics [17]. - In human evaluations, the excellent rate for realism increased from 8.2% to 38.9% and for aesthetic quality from 9.8% to 40.5% after SRPO training [17][18]. - Notably, a mere 10 minutes of SRPO training allowed FLUX.1-dev to surpass the latest open-source version FLUX.1.Krea on the HPDv2 benchmark [19].