Sketches offer a simple yet powerful way to represent object configurations, making them ideal for local image structure manipulation. Traditional methods often treat sketch-based editing as an image inpainting task, requiring both user-provided strokes and masks, which hinders the user experience. Although recent mask-free stroke-based editing methods are more convenient, they often produce significant artifacts or unintentionally modify irrelevant regions. To overcome these challenges, we propose DiffStroke, a mask-free method for high-quality image editing using only partial sketches. Trainable plug-and-play Image-Stroke Fusion (ISF) modules and an effective mask estimator are developed to address the limitations of previous conditional control diffusion models in preserving style consistency and protecting irrelevant areas. The ISF modules fuse stroke encodings with source image features as input conditions, enabling DiffStroke to control local shapes while preserving overall style consistency. The mask estimator automatically predicts masks to preserve irrelevant regions without the need for manual input. Specifically, DiffStroke blends the estimated clean latent image with the encoded source image using the predicted mask, with the mask estimator trained to minimize the error between the blended result and the latent target image. Experimental results on natural and facial images demonstrate that DiffStroke outperforms previous methods in both simple and complex stroke-based image editing tasks.
Keywords: Image manipulation, sketch-based image editing, mask-free, diffusion model
TL;DR: A conditional control diffusion model for high-quaility, mask-free image manipulation with partial sketches.
Abstract:
Supplementary Material: zip
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/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 6597
Loading