Text-Driven Image Editing using Cycle-Consistency-Driven Metric Learning
Keywords: Diffusion Models, Text-Driven Image-to-Image Translation
TL;DR: We propose effective guidance terms using the representations given by the pre-trained CLIP and Stable Diffusion while employing the cycle-consistency objective for diffusion models.
Abstract: We present a simple but effective training-free method for text-driven image-to-image translation based on pretrained text-to-image diffusion models. Since a naive application of the pre-trained diffusion models for the manipulation tasks often significantly destroys the structure or background of the source image, we revise the original backward process for the target image by meaningfully aligning better with a given target task while preserving the background or structure of a source image. We derive a new guidance objective term that is a combination of maximizing the similarity with target prompts rather than the source prompt based on the pre-trained CLIP and minimizing the distance with the source latents. Moreover, contrary to existing methods based on the diffusion models, we exploit the cycle-consistency objective in order to further maintain the background of the source image, where we perform an iterative optimization process by alternately optimizing the source and target latents. Experimental results demonstrate that the proposed method achieves outstanding editing performance on various tasks when combined with the pre-trained Stable Diffusion.
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 669
Loading