EditHOI: A framework for HOI image editing with self-generated skeleton guidance

ChangHee Yang, Chan Hee Kang, Kyeongbo Kong, Hanni Oh, Suk-Ju Kang

Keywords: Image editing, object interaction human skeleton generation

Abstract:

Recently, there were remarkable advances in image editing tasks. This could be categorized into text-guided global editing, local editing, text-guided local editing. To resolve the flawed human generation problem in prior image editing models, we propose a novel skeleton and text-guided local editing framework, EditHOI. Our goal is to edit an image by synthesizing an object-interactive human in the image. To do this, our framework consists of two stages: the first stage generates object-interactive skeleton using diffusion-based module, while the second stage outputs a Human and Object Interaction (HOI) image based on skeleton and text guidance. For effective evaluation on a object-interactive skeleton, we designed joint parameter and two evaluation metrics; object interaction top-n accuracy and skeleton probability distance. The excellent performance of our framework is demonstrated through experiments qualitatively and quantitatively. Lastly, we show its applicability such as user controllable editing, generating pseudo SMPL ground truth and scalability to human-to-human interaction. The corresponding code is available at https://anonymous.4open.science/r/ HOI_editing_image-43F1/

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: 2394