Rapid Iterative Inference for Text-Guided Image Editing

ABSTRACT

With the development of deep learning, image editing has transformed from traditional manual methods to text-based techniques. Traditional methods are inefficient and require high professional skills, whereas text-based editing allows users to guide the process through natural language, improving convenience and operability. Diffusion models have shown significant progress in image generation and editing, but they typically need many iterative steps to produce high-quality images, leading to slow speeds. To address this, we propose a text-guided fast iterative inference image editing method. This approach applies the DDPM noise inversion method to the Adversarial Diffusion Distillation model. However, the basic framework often results in visual artifacts and weak editing strength. To improve visual effects, we introduce a time-step offset method. To enhance editing accuracy, pseudo-guidance and semantic guidance are added, ensuring generated images align better with text prompts. Combining these strategies, we design a collaborative optimization algorithm that achieves fast editing through noise inversion, time-step offset, and guidance techniques. Experiments show that our method achieves semantic changes consistent with text prompts while maintaining image structure. It completes editing in 0.806 seconds, suitable for real-time interaction. Our method outperforms existing few-step sampling models and matches multi-step models in quality.