SRPO - A text-based image model launched by Tencent Hunyuan

What is SRPO?

SRPO (Semantic Relative Preference Optimization) is a text-to-image generation model developed by Tencent Hunyuan. By designing the reward signal as a text-conditional signal, it enables online adjustment of the reward, reducing reliance on offline reward fine-tuning. SRPO introduces Direct-Align technology, directly recovering the original image from any time step through predefined noise priors, avoiding over-optimization in later time steps. Experiments on the FLUX.1.dev model show that SRPO significantly improves the realism and aesthetic quality of generated images as perceived by humans, and its training efficiency is extremely high, completing optimization in just 10 minutes.

SRPO's main functions

• Improve image generation qualityBy optimizing the diffusion model, the generated images are significantly improved in terms of realism, detail richness, and aesthetic quality.
• Online reward adjustmentIt allows users to dynamically adjust reward signals via text prompts, changing the style and preferences of image generation in real time without the need for offline fine-tuning of the reward model.
• Enhance the adaptability of the modelThis enables diffusion models to better adapt to different task requirements and human preferences, such as optimization at different lighting conditions, styles, or levels of detail.
• Improve training efficiencyBy optimizing the early stages of the diffusion process, SRPO can complete model training and optimization in a short time (e.g., 10 minutes), significantly improving training efficiency.

SRPO technical principles

• Direct-Align technology:During training, SRPO injects Gaussian noise into clean images and recovers the original image through a single-step denoising operation.In this way, SRPO can effectively avoid over-optimization problems that occur in the later time steps of the diffusion process and reduce reward hacking (such as the model using the bias of the reward model to generate low-quality images).Compared to traditional methods, SRPO can be optimized at earlier time steps, which helps to improve training efficiency and generation quality.
• Semantic Relative Preference Optimization (SRPO):The reward signal is designed as a text conditional signal, and the reward signal is adjusted by positive and negative cue words.The model is optimized by calculating the reward difference between positive and negative cue word pairs.SRPO supports dynamic adjustment of reward signals during training, enabling the model to adjust the generation strategy in real time according to different task requirements.
• Reward Aggregation Framework:To improve the stability of the optimization, SRPO injects noise multiple times during the training process to generate a series of intermediate images, and performs denoising and restoration operations on each image.By aggregating intermediate rewards using a decay discount factor, SRPO effectively reduces reward hacking in later time steps, improving the overall quality of the generated image.

SRPO project address

• Project official websitehttps://tencent.github.io/srpo-project-page/
• GitHub repositoryhttps://github.com/Tencent-Hunyuan/SRPO
• HuggingFace model libraryhttps://huggingface.co/tencent/SRPO
• arXiv technical paper: https://arxiv.org/pdf/2509.06942v2

SRPO application scenarios

• Digital art creationArtists and designers can generate high-quality digital artworks, dynamically adjust image styles through text prompts, and achieve rapid iteration from concept sketches to final works.
• Advertising and MarketingAdvertising agencies can generate images that match specific brand styles and market positioning, quickly producing a variety of design options and improving creative efficiency.
• Game developmentGame developers generate high-quality game textures, character designs, and scene backgrounds, enhancing the game's visual effects and player experience.
• Film and television productionIn film and television production, it is used to generate realistic special effects scenes, backgrounds, and characters, reducing post-production time and costs.
• Virtual Reality (VR) and Augmented Reality (AR)The model can generate high-quality virtual environments and objects, enhancing the immersion and realism of VR and AR applications.