Neural Contrast: Leveraging Generative Editing for Graphic Design Recommendations

TL;DR

A generative approach using a diffusion model is proposed, which ensures the altered regions beneath design assets exhibit low saliency while enhancing contrast, thereby improving the visibility of the design asset.

Abstract

Creating visually appealing composites requires optimizing both text and background for compatibility. Previous methods have focused on simple design strategies, such as changing text color or adding background shapes for contrast. These approaches are often destructive, altering text color or partially obstructing the background image. Another method involves placing design elements in non-salient and contrasting regions, but this isn't always effective, especially with patterned backgrounds. To address these challenges, we propose a generative approach using a diffusion model. This method ensures the altered regions beneath design assets exhibit low saliency while enhancing contrast, thereby improving the visibility of the design asset.

Figures (5)

• Figure 1: Our generative editing pipeline enhances contrast in graphic design elements, such as text or SVGs, on background images. This pipeline automatically adjusts the luminance to contrast with text color, thereby emphasizing the text. As shown in examples 1, 4, and 6, luminance changes in the opposite direction of text color to enhance contrast. Examples 2 and 3 demonstrate text emphasis through the creation of contrasting elements. Additionally, contrast is further improved by reducing image energy, such as removing background objects behind the text, as seen in examples 4 and 5. This automated process involves injecting color, luminance, content, and fractal noise, followed by generative editing with a diffusion model.
• Figure 2: An overview of the design editing/generative pipeline, detailing all intermediary stages and outcomes, except for Design Variations. The input prompt "autumn in the forest, close up, macro, morning light, wallpaper, leaves blanket, multiple leaves" forms the input of the pipeline. Dashed lines separate distinct steps, with Design Variation as the exception. Diamond symbols denote algorithms, while rectangular shapes represent models. The pipeline input comprises the layout combined with either the $P$ prompt alone, the design background alone, or both.
• Figure 3: The plot of the function $\mathcal{F}$ post-training using the dataset pairs $X$ for input data and $Y$ for corresponding labels. Left: $\mathcal{F}$ function for the model $\omega=$ SDXL-base podell2023sdxl model. Right for the model $\omega^*=$ SD-v1.5 rombach2022high
• Figure 4: Design Variation Potential: The first column shows the input prompt. The second column displays the initial design, marked with a red border for visibility at reduced readability. The next two columns present variations of the original layout, followed by the final two columns which feature new layout-generated variations.
• Figure 5: The investigation focuses on evaluating the influence of the strength parameter on the SDXL-base model podell2023sdxl within a subset of the utilized data sourced from Crello 8yamaguchi2021canvasvae. The objective is to assess the impact of the strength parameter on the final outcomes of each model, where $\omega\left(\cdot, \cdot, \cdot\right)$ is the Diffusion Generative Network in the DiffEdit 13couairon2022diffedit or SDEdit paradigm 14meng2021sdedit, where the first parameter is the constant prompt $P$ for all generated images, the second parameter is the constant starting image $I$ (obtained precisely as in Subsection \ref{['AH']}) for all generated images and the third parameter is the intensity that is desired from $I$, or the strength parameter from SDEdit and DiffEdit.