InstantStyle is a tuning-free framework for style-preserving text-to-image generation. It addresses the challenges of style consistency, style degradation, and content leakage in image generation. The framework introduces two key strategies: 1) Decoupling style and content from reference images in the feature space by subtracting content text features from image features, which effectively reduces content leakage. 2) Injecting reference image features only into style-specific blocks, preventing style leaks and eliminating the need for complex weight tuning. These strategies enable effective style transfer without the need for paired datasets or additional modules. InstantStyle achieves superior visual stylization, balancing style intensity and text controllability. It is model-independent, pluggable, and compatible with other attention-based feature injection methods. The framework is implemented on Stable Diffusion XL (SDXL) and demonstrates strong performance in style transfer tasks, achieving high consistency across various content and styles. The method is validated through extensive experiments, showing its effectiveness in reducing content leakage and maintaining style integrity. InstantStyle has the potential to significantly impact downstream tasks and other domains by enabling consistent generation with preserved style.InstantStyle is a tuning-free framework for style-preserving text-to-image generation. It addresses the challenges of style consistency, style degradation, and content leakage in image generation. The framework introduces two key strategies: 1) Decoupling style and content from reference images in the feature space by subtracting content text features from image features, which effectively reduces content leakage. 2) Injecting reference image features only into style-specific blocks, preventing style leaks and eliminating the need for complex weight tuning. These strategies enable effective style transfer without the need for paired datasets or additional modules. InstantStyle achieves superior visual stylization, balancing style intensity and text controllability. It is model-independent, pluggable, and compatible with other attention-based feature injection methods. The framework is implemented on Stable Diffusion XL (SDXL) and demonstrates strong performance in style transfer tasks, achieving high consistency across various content and styles. The method is validated through extensive experiments, showing its effectiveness in reducing content leakage and maintaining style integrity. InstantStyle has the potential to significantly impact downstream tasks and other domains by enabling consistent generation with preserved style.