Efficient Enhancement to Diffusion-Based Image Editing

AIGC has become a fascinating domain for research and application with the proposing of advanced generated models. Among all of the methods, diffusion probabilistic model exhibits excellent ability in both the diversity and fidelity of the sample. And text2img model can be constructed by training this type of model with multi-modal data, which makes it available for users to generate the image according to text prompt. Albeit text2img model is generalized to distinct scenarios, this also causes the problem for customized image generation. And it is exactly what the down-stream task called image editing aims to solve. There are various methods to conduct editing to the source image. But most of them concern about the specific manipulations, which are based on the cross-attention mechanism to deal with the correlation between the text and image. The naive sampling process is utilized and the effect of the sampling strategy to the editing result is not investigated. As the sampling procedure of image editing is also an iterative denoising process of the diffusion model. Practical sampling strategy is equally critical to the editing result. We introduce an efficient enhancement namely dynamic guidance dedicated for image editing to customize the interval applying the classifier-free guidance. This sampling strategy can not only enhance the quality of the editing image but also accelerate the inference efficiency.