Yucheng Yang, Xiaoxiao Sun, Shusheng Yang, Jiakang Chen, Serena Yeung-Levy
This paper introduces Semantic Feedback Guidance (SFG) to improve multimodal diffusion transformers (MMDiTs) by suppressing image-to-text feedback. SFG guides the model away from preserving errors in the evolving image. This leads to improved generation quality and alignment with the prompt.
The proposed method has practical significance as it enhances the performance of image generation models, which can be applied to various applications such as art, design, and data augmentation. Improved image generation quality can also benefit downstream tasks like object detection and image segmentation. This can lead to more realistic and accurate results in these applications.
Thank you to arXiv for use of its open access interoperability. arxiv.org. aibrief.ai is not affiliated with or endorsed by arXiv.