SRPO (Semantic Relative Preference Optimization)

SRPO is an optimization method developed by the Tencent Hunyuan team for text-to-image generation tasks.

Results: Experiments on the FLUX.1 - dev model demonstrate that SRPO significantly enhances the realism and aesthetic quality of generated images in human evaluations. The original FLUX model had an excellence rate of only 8.2% for realism, which surged to 38.9% after SRPO training. The aesthetic quality excellence rate increased from 9.8% to 40.5%, with the overall preference reaching an excellence rate of 29.4%.

---------------------------------

LoRAs extracted from the SRPO Flux.1-Dev FP8 model aim to provide modular and lightweight adaptability without requiring a fully fine-tuned model. They support flexible blending, reduce storage costs, and enable efficient experimentation.

---------------------------------

SRPO Workflow Link:

https://www.liblib.art/modelinfo/97c008dcca214e7ca69077b484e740b7

---------------------------------

Key Features of SRPO:

Enhanced Image Generation Quality: Fine-tunes diffusion models to deliver noticeable enhancements in detail rendering, visual realism, and artistic aesthetics in output images.

Dynamic Reward Adjustment Support: Users can adjust reward guidance in real time by inputting positive and negative text prompts, flexibly controlling image style and content preferences without retraining or fine-tuning the reward model.

Improved Model Generalization: Enables the model to quickly adapt to diverse human aesthetics and task requirements, such as generation goals across different lighting, artistic styles, or detail levels.

Efficient Training Mechanism: Focuses optimization on the early stages of the diffusion process, enabling model tuning to be completed in extremely short timeframes—such as within 10 minutes—greatly improving iteration speed and resource utilization.

---------------------------------

Core Technical Principles of SRPO

Direct - Align Technology: By pre-injecting noise and leveraging a preset noise prior, the original image can be restored from any timestep. This avoids the limitations of optimizing solely in late steps, reduces "reward hacking," and alleviates the gradient explosion issue faced by traditional methods during backpropagation at early timesteps.

Semantic Relative Preference Optimization: Models rewards as difference signals guided by positive and negative text prompts. For the same image, the model calculates rewards using positive and negative prompts respectively, then takes their relative difference as the optimization target to achieve real-time control over the generation process.

Reprinted from: Flux.1-Dev SRPO Realism & Aesthetic Enhancement