Tencent Releases SRPO Open-Weights Image Generation Model

Introduction

On September 1, 2025, Tencent expanded its sophisticated generative AI portfolio with the release of SRPO (Semantic Relative Preference Optimization). This new research model, available as open weights on Hugging Face, represents a significant step forward in how diffusion models are aligned with complex human aesthetic and semantic preferences.

While Tencent's HunyuanImage line continues to drive its enterprise-grade visual generation, SRPO serves as a specialized research vehicle. It aims to solve a fundamental problem in generative modeling: the gap between simple text-to-image alignment and the nuanced, fine-grained preferences that humans use to judge visual quality and compositional accuracy.

• Release Date: September 1, 2025
• Developer: Tencent (Hunyuan Team)
• Availability: Open-weights via Hugging Face
• Core Innovation: Semantic Relative Preference Optimization

Key Features & Architecture

The technical core of SRPO lies in its departure from standard reward modeling. Instead of treating image generation as a single-step optimization, SRPO focuses on 'Directly Aligning the Full Diffusion Trajectory.' This means the model optimizes the entire denoising process rather than just the final output.

By implementing Semantic Relative Preference Optimization, the model uses text-conditioned signals to formulate rewards. This allows the model to understand not just 'what' an object is, but 'how' it should be positioned and styled relative to the semantic nuances provided in the prompt. As a fine-tune of the FLUX.1 [dev] architecture, it leverages a high-parameter foundation to achieve superior structural integrity.

• Architecture: Fine-tuned FLUX.1 [dev] base
• Optimization Method: Full Diffusion Trajectory Alignment
• Reward Mechanism: Text-conditioned semantic signals
• Quantization Support: Available in 8-bit (fp8_e4m3fn/Q8_0) versions

Performance & Benchmarks

SRPO has demonstrated exceptional capabilities in aligning visual outputs with complex linguistic instructions. In comparative testing, the model achieved an Artificial Analysis Elo score of 831, placing it among the top-tier research models for prompt adherence and aesthetic quality.

The model's evaluation methodology focuses on fine-grained human preference. Unlike traditional benchmarks that might only measure CLIP scores, SRPO is evaluated on its ability to navigate the 'diffusion trajectory'—ensuring that every step of the denoising process moves closer to the semantic intent of the user. This results in fewer compositional errors and higher fidelity to long-form descriptive prompts.