SDE-RAE:CLIP-based realistic image reconstruction and editing network using stochastic differential diffusion

Generative Adversarial Networks (GANs) has long dominated the field of image reconstruction and editing. It is capable to train a generator in an adversarial way, which can fool the discriminator and enable the generated image to be of high quality. However, this approach is often difficult to train, and the final result is hard to converge. Each different style of image requires construction of different datasets and complex optimization functions, and the training process is uncertain. To solve this problem, we propose a realistic image reconstruction and editing method based on Stochastic Differential Equation (SDE-RAE), where the diffusion model converts Gaussian noise to real photos by iterative denoising. What we only need to do is to construct simple loss functions in the reconstruction process to achieve high-quality image reconstruction, and we propose a novel semantic enhancement CLIP (Contrastive Language-Image Pre-Training) to interfere with the SDE parameter optimization direction in the editing process. Simple text is needed to achieve unique image editing. Our method generates high-quality images that retain the texture and contour features of the original image. Specifically, we manipulate the initial image, perturb the image by adding random noise, and then iteratively denoise the image by reverse SDE, manipulating the image's RGB pixels to achieve image reconstruction and editing. Code and dataset https://github.com/haizhu12/SDE-RAE.

Automated summary

Generative Adversarial Networks (GANs) have been widely used in image reconstruction and editing. However, they are difficult to train and converge. To overcome this challenge, we propose a realistic image reconstruction and editing method based on Stochastic Differential Equation (SDE-RAE), which achieves high-quality image reconstruction with simple loss functions and interferes with parameter optimization using semantic enhancement.