Automated image generation

Abstract

Multi-conditional image generation aims to synthesize images that satisfy diverse conditions, such as textual descriptions, segmentation masks, and landmark constraints. Current training-free approaches, which rely on off-the-shelf and open-source pre-trained networks to provide guidance, perform well for single conditions but fail to capture the complex interdependencies among multiple conditions. The purpose of this project is to develop a novel framework that overcomes these limitations by effectively modelling the interactions between conditions. To achieve this, the project analyzes existing methods and introduces an innovative design: a time-independent approximated energy guidance function enhanced with interaction modelling. This function captures non-linear and complex dependencies, guiding an iterative denoising process to progressively refine the generated images. Experimental results indicate that our approach outperforms existing techniques, producing images that are both coherent and condition-consistent. In conclusion, the framework not only resolves key challenges in multi-conditional image synthesis but also provides a basis for future research, with recommendations to further explore adaptive energy functions for even broader applicability.