Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
102 tokens/sec
GPT-4o
59 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
50 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Attentive Eraser: Unleashing Diffusion Model's Object Removal Potential via Self-Attention Redirection Guidance (2412.12974v3)

Published 17 Dec 2024 in cs.CV

Abstract: Recently, diffusion models have emerged as promising newcomers in the field of generative models, shining brightly in image generation. However, when employed for object removal tasks, they still encounter issues such as generating random artifacts and the incapacity to repaint foreground object areas with appropriate content after removal. To tackle these problems, we propose Attentive Eraser, a tuning-free method to empower pre-trained diffusion models for stable and effective object removal. Firstly, in light of the observation that the self-attention maps influence the structure and shape details of the generated images, we propose Attention Activation and Suppression (ASS), which re-engineers the self-attention mechanism within the pre-trained diffusion models based on the given mask, thereby prioritizing the background over the foreground object during the reverse generation process. Moreover, we introduce Self-Attention Redirection Guidance (SARG), which utilizes the self-attention redirected by ASS to guide the generation process, effectively removing foreground objects within the mask while simultaneously generating content that is both plausible and coherent. Experiments demonstrate the stability and effectiveness of Attentive Eraser in object removal across a variety of pre-trained diffusion models, outperforming even training-based methods. Furthermore, Attentive Eraser can be implemented in various diffusion model architectures and checkpoints, enabling excellent scalability. Code is available at https://github.com/Anonym0u3/AttentiveEraser.

Overview of AAAI Press Formatting Instructions for LaTeX Authors

The document titled "AAAI Press Formatting Instructions for Authors Using LaTeX — A Guide" provides comprehensive guidelines for preparing manuscripts intended for submission to conferences or publications facilitated by AAAI Press. This guide serves as a vital resource for authors to ensure uniformity and adherence to standardized formatting protocols when using LaTeX, which is a widespread document preparation system in academic publishing, particularly in fields like computer science.

Core Formatting Requirements

The paper emphasizes several essential requirements for manuscript preparation:

  • Use of AAAI Style Files: Authors must strictly use the provided 2025 AAAI Press LaTeX style files (aaai25.sty) and the corresponding bibliography style files (aaai25.bst). These files ensure that the documents adhere to specific visual and structural standards.
  • PDF Compilation: It is imperative that the final output is a PDF compiled using PDFLaTeX, with all fonts embedded, thereby ensuring fidelity across different viewing platforms and devices. Notably, the use of PostScript (.ps) and Encapsulated PostScript (.eps) files for figures is discouraged in favor of .jpg, .png, or .pdf.
  • Document Integrity and Readability: The paper specifies constraints on typographic alterations—such as spacing, font sizes, line spacing, and document layout alterations—prohibiting modifications that could disrupt the uniform appearance essential for scholarly publications. Additionally, type 3 fonts, which adversely affect print quality, are not permitted.
  • Manuscript Structure: Authors are provided with guidance on document segmentation from introduction to references, emphasizing a two-column format on US letter-size paper, with precise margin and column dimensions. The paper must maintain a consistent 10-point font size in Times Roman or Nimbus.

Submission and Compliance Processes

The document delineates a meticulous checklist for submission to ensure the manuscript is publication-ready:

  • File Submission: Authors must submit their LaTeX source file along with a fully compliant PDF, any necessary bibliography (.bib) files, and all pertinent graphics. The guidelines underscore that submissions must be concise and free of unused files.
  • Electronic Submission and Naming Convention: Specific protocols are outlined for electronic file submission, including a naming convention based on the surname of the first author to streamline the submission process.
  • Copyright Requirements: Submissions must include signed copyright forms, and the appropriate copyright notices must be visibly incorporated in the document.

Prohibited Practices and Common Pitfalls

Authors are cautioned against the use of certain commands, macros, and packages that could disrupt the document's format, as outlined in comprehensive tables within the document. Additionally, the guide identifies common pitfalls in formatting that could lead to rejection or the need for resubmission, such as improper font inclusion, inadequate figure resolution, and undesirable styling adjustments.

Implications and Future Directions

This document is integral for maintaining the quality and cohesion of publications within AAAI Press proceedings, reflecting broader standards in academic publishing that prioritize consistency, accessibility, and readability. As the field of artificial intelligence continues to expand, adherence to such standardized guidelines ensures that scholarly communication remains clear and professional. The emphasis on accessibility, particularly through restrictions on color use, suggests a forward-looking approach considering the needs of a diverse readership, including those with visual impairments.

In the evolving landscape of AI research, the implications of these guidelines permeate not only the format but also the dissemination and impact of academic work. Adopting and adhering to these standards facilitates the engagement of researchers across different domains, contributing to a cohesive and collaborative academic community.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (4)
  1. Wenhao Sun (31 papers)
  2. Benlei Cui (3 papers)
  3. Jingqun Tang (22 papers)
  4. Xue-Mei Dong (1 paper)
Citations (2)
Reddit Logo Streamline Icon: https://streamlinehq.com