SRPO - Text-to-Image Generation Model launched by Tencent Mixed Meta

What is SRPO

SRPO (Semantic Relative Preference Optimization) is a text-to-image generation model introduced by Tencent Mixed Meta, which optimizes the reward mechanism through text conditioned signals to achieve online adjustment of rewards and reduce offline fine-tuning dependency.SRPO introduces Direct-Align technology to avoid over-optimization at a later stage and improve training efficiency. The model can significantly improve the realism and aesthetic quality of the generated images, and is widely used in digital art creation, advertising and marketing, game development, film and television production, and VR/AR, providing creators with efficient and flexible image generation solutions.

Features of SRPO

• Image Quality Improvement: By optimizing the diffusion model, SRPO is able to generate more realistic and detailed images, significantly improving the realism and aesthetic quality of the images.
• Dynamic reward adjustments: Supports users to adjust reward signals in real time based on text prompts, eliminating the need for offline fine-tuning and enabling dynamic changes in image styles and preferences.
• Increased adaptability: The model is better adapted to different task requirements, such as optimization for different lighting conditions, styles or levels of detail, and is highly flexible.
• Efficient training: By optimizing the early stages of the diffusion process, SRPO can complete training and optimization in a short period of time, greatly improving training efficiency and saving time and resources.

SRPO's core strengths

• Online Reward Adjustment: Dynamically adjusting reward signaling through positive and negative cue words reduces reliance on offline reward fine-tuning and improves model flexibility.
• Improve the quality of image generation: The model optimizes the early time step of the diffusion model to significantly enhance the realism, detail and aesthetic quality of the image.
• Avoid rewarding hacking: Effectively suppressing reward hacking and enhancing training stability with relative preference mechanisms and negative reward signals.
• Flexibility and scalability: Text-based conditional signaling, with simple text prompts to adjust the image style to suit a wide range of tasks.

What is SRPO's official website?

• Project website:: https://tencent.github.io/srpo-project-page/
• GitHub repository:: https://github.com/Tencent-Hunyuan/SRPO
• HuggingFace Model Library:: https://huggingface.co/tencent/SRPO
• arXiv Technical Paper:: https://arxiv.org/pdf/2509.06942v2

Who SRPO is for

• Digital artists and designers: Rapidly generate and iterate high-quality digital artworks with the help of models, flexibly adjust image styles through text prompts, and achieve efficient visualization of creativity.
• Advertising and marketing staff: Use models to generate images that match brand styles, quickly produce multiple design options, improve creative efficiency, and reduce design costs.
• game developer: Accelerate the development process and enhance game visuals by generating high-quality game textures, character and scene backgrounds.
• moviemaker: Use models to generate realistic special effects scenes and characters, reduce post-production costs, and improve the visual quality of film and television productions.
• VR and AR Developers: Rely on modeling into high-quality virtual environments and objects to enhance the immersion and realism of VR and AR applications.