Scholarlink

<p>In this paper, we present an image editing model capable of precisely generating and modifying objects within both real and synthetic images. Our model operates in two steps. The first step involves editing and generating new content in a user-specified area, using a zero-shot approach with multimodal guidance that combines textual descriptions and a 2D geometric shape to direct the editing process. This approach leverages both the semantic information provided by the text and the visual details of the image, resulting in more precise and consistent modifications while preserving the visual integrity of the original image. The second step integrates an illumination adjustment mechanism. This step ensures that the lighting in the modified area remains consistent with the rest of the image, thus avoiding visual artifacts and maintaining high aesthetic quality. Our experiments were conducted on both synthetic and real datasets, demonstrating the superior performance of our model in terms of visual quality, structural coherence, and lighting adjustment. Our approach minimizes the artifacts often encountered in other methods and offers great flexibility in various image editing scenarios. For more details, visit our project website https://multilightdiff.github.io/.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

MultiLightDiff: Multimodal Diffusion Approach for Precise Image Editing and Illumination Adjustment

  • 期刊论文

  • 发表日期：

    2026年5月19日

查看全文
  • Fayçal Abbas,
  • Mehdi Malah,
  • Ramzi Agaba,
  • Mohamed Chaouki Babahenini

摘要

In this paper, we present an image editing model capable of precisely generating and modifying objects within both real and synthetic images. Our model operates in two steps. The first step involves editing and generating new content in a user-specified area, using a zero-shot approach with multimodal guidance that combines textual descriptions and a 2D geometric shape to direct the editing process. This approach leverages both the semantic information provided by the text and the visual details of the image, resulting in more precise and consistent modifications while preserving the visual integrity of the original image. The second step integrates an illumination adjustment mechanism. This step ensures that the lighting in the modified area remains consistent with the rest of the image, thus avoiding visual artifacts and maintaining high aesthetic quality. Our experiments were conducted on both synthetic and real datasets, demonstrating the superior performance of our model in terms of visual quality, structural coherence, and lighting adjustment. Our approach minimizes the artifacts often encountered in other methods and offers great flexibility in various image editing scenarios. For more details, visit our project website https://multilightdiff.github.io/.