DE-net: Dynamic Text-Guided Image Editing Adversarial Networks

Text-guided image editing models have shown remarkable results. However, there remain two problems. First, they employ fixed manipulation modules for various requirements (e.g., color changing, texture content adding and removing), which results in over-editing or insufficient editing. Second, do not clearly distinguish between text-required text-irrelevant parts, leads to inaccurate To solve these limitations, we propose: (i) a Dynamic Editing Block (DEBlock) that composes different dynamically requirements. (ii) Composition Predictor (Comp-Pred), predicts the composition weights DEBlock according inference on target texts source images. (iii) text-adaptive Convolution (DCBlock) queries features parts parts. Extensive experiments demonstrate our DE-Net achieves excellent performance manipulates images more correctly accurately.