How are we using generative AI for images?

From the AI and data department of hiberus, we employ generative AI for a wide range of applications, having the privilege of being at the forefront in the implementation and experimentation of emerging technologies. Generative Artificial Intelligence is revolutionizing various industries, especially highlighting its impact in the retail and marketing sector.

In this article, I will share how at hiberus we are leveraging generative AI to drive innovations in marketing campaigns, fashion design, and enhancing customer experience through cutting-edge technologies. 

 

How to generate images using AI

1. Automatic generation of creatives for social networks (Advertising banners) 

Recently, at hiberus, we have undertaken one of the most interesting and innovative applications: the use of generative AI for the automatic creation of advertising banners intended to promote the subscription to the Heraldo digital newspaper. This proof of concept has proven to be highly effective and valuable by completely revolutionizing the design and distribution process of creatives on social networks like Instagram, Facebook, and Twitter. 

 

 

The Challenge 

Previously, the creation of advertising banners involved a manual process that required a considerable amount of time and resources from the design team. The need to produce variations of creatives for different platforms and sizes added complexity and delays to the process. Moreover, consistency in style, colors, and typography fonts was a constant challenge due to the variability inherent in manual work. 

 

The Solution: Automatic Generation of Creatives 

To address these challenges, we developed a web application that uses generative AI to automate the creation of advertising banners. The application allows users to enter a descriptive ticket of the campaign, select specific options for variations in styles, and obtain the creatives as PNG images, all tailored to the specific dimensions, styles, colors, and typography fonts required by Heraldo. 

 

How it works 

The process begins when the user enters a descriptive ticket of the campaign into the application. This description includes details about the campaign’s message, the target audience, and any style or branding specifications that must be considered. Next, the user selects the specific dimensions for the social media platforms where the campaign will be promoted and adjusts the style options as needed.  

The AI processes this information and generates a series of advertising banners that are not only visually appealing but also consistent with Heraldo’s brand identity. The generated banners are optimized for each social media platform, ensuring proper display on any device or screen resolution. 

 

 

Results and Benefits 

This use case has multiple benefits: 

• Improved efficiency: The automatic generation of banners has drastically reduced production time, enabling a more agile launch of campaigns. 
• Brand consistency: Automation ensures flawless consistency in all design aspects, reinforcing Heraldo’s brand identity across all platforms. 
• Personalization at scale: The ability to effortlessly generate personalized variations for different audience segments or platforms. 
• Innovation: This project highlights Hiberus’s commitment to innovation and the application of emerging technologies to solve real business problems. 

 

2. Mixing clothing garments and styles 

Continuing our push towards innovation in the use of Generative Artificial Intelligence, at hiberus we have taken creativity in fashion design to a new level. The latest proof of concept we have implemented uses AI to combine styles from two different images, whether they are clothing garments or other products, resulting in creations that are as unique as they are fascinating. This project stands out not only for its ability to harmoniously blend styles but also for the technical development behind the final user experience. 

 

The challenge of combining styles 

Fashion is a domain where originality and customization are key. However, the challenge lies in how to effectively merge styles without losing the essence of the original pieces. Traditionally, this process would require considerable effort from designers to conceptualize and execute combinations that are both aesthetically pleasing and functionally viable. 

 

Innovative Solution: AI Pipeline for Garment Mixes 

To overcome these challenges, we have developed a backend and frontend app demo pipeline that allows users to create combinations of clothing garments using AI. This pipeline includes several critical steps to ensure that the final results are high quality: 

• Detailed description and tagging of images: We use vision-language models (ViT) to describe images and their styles in detail, adapting to the specific context of the customer’s industry. In this way, we detect important characteristics, for example, “Art Deco,” “Coco Chanel,” “Burgundy color,” “2011 trend,” or “Mahogany wood”. 
• Style combination: We use advanced generative AI algorithms to combine the visual elements of two input images, creating a hybrid design that retains key features of both.  
• Hand detection and inpainting: We recognize that certain details, like hands, can be challenging to replicate accurately for AI. Therefore, we have integrated hand detection using another AI model, followed by an inpainting process to realistically and consistently repaint these areas with the new design. 
• Customization through the web app: Our web application allows users to specify detailed generation options, such as the input image priority, specific colors to highlight, among others. This ensures that the final result is aligned with the user’s personal preferences. 

 

User Experience 

The user interface of the app demo is intuitive, allowing users to easily upload the images of the garments they wish to combine and adjust their design preferences. Once the user finalizes their selection, the system processes the images through the pipeline and presents a series of combined designs for review.

This approach not only simplifies the creative process for designers, allowing them to experiment with bold and unique combinations, but also offers our clients the opportunity to explore and generate exclusive garments. 

 

3. Virtual Try-On for Various Products 

The culmination of our efforts in generative AI is the development of virtual try-on technologies. This allows customers to virtually try on products for their home, as well as clothing items, through modification and generation of images with AI, seeing themselves in different outfits in a variety of situations.

It not only enhances the shopping experience but also significantly reduces return rates by providing a more accurate representation of how the garment will look on the customer.

 

 

Revolutionizing Product Visualization 

Our technology strategy focuses on combining advanced artificial intelligence techniques to analyze and understand the customer’s space. Through a sophisticated process, we are able to intelligently segment the user’s space images, identifying key areas and virtually adapting the selected products to their environment. This approach allows users to get a realistic preview of how furniture, accessories, or any other home product would look in their space. 

 

 

Personalization and Precision 

The heart of our project is the use of advanced algorithms that not only understand the space and dimensions of rooms but also adjust the products in the right context, taking into account lighting, perspective, and aesthetic harmony. This precision ensures that the virtual representation is as close as possible to how the product would look in reality. 

 

 

Innovation in the Shopping Experience 

To achieve a convincing and high-quality visual integration, we apply image enhancement techniques that include fine-tuning models specific to products and environments, as well as optimizing image resolution to ensure that product details are clear and precise. Our focus is on preserving the essence of the user’s home while offering personalized recommendations that fit perfectly into their space and personal style. 

 

Ease and Accessibility 

This project is not only designed to improve the online shopping experience but also to make it more accessible. Users can experiment with different product options without the need to visit a physical store or imagine how the product would look in their space. This tool is valuable not only for customers but also benefits retailers by reducing return rates and increasing customer satisfaction.

In summary, the virtual try-on of home products and furniture is a step forward in personalizing the online shopping experience. By using advanced technologies, we are bridging the gap between imagination and reality, enabling customers to design and personalize their spaces with confidence and precision. This project underscores our ongoing commitment to innovation and our desire to provide exceptional experiences to our users. 

 

Maintaining Consistency  

As we delve further into the application of Generative Artificial Intelligence within hiberus, we face the challenge of maintaining consistency in the results, especially in complex scenarios such as product visualization in various poses or contexts. 

 

The Role of Image Processing  

ControlNet is an image processing framework and vision models that we have integrated into our image generation process, allowing us to maintain consistency in the pose and shape of garments across different images. The tool is especially valuable when we need to visualize how the same garment would look on different bodies or in a variety of poses, a common task in the fashion industry and online retai. 

 

ControlNet for Consistency in Images 

One of the uses we implement with ControlNet is for mapping poses and the body structure of the input to a representation that AI can understand and replicate. This process involves several key steps: 

• Pose Detection: Initially, the pose of the subject in the input image is detected, identifying key body points. 
• Control Map generation: From this detection, ControlNet generates a “control map” that represents the pose and body proportions. This map guides the AI to maintain consistency in the pose across different images. 
• Image Generation with Control Map: Using the control map as a reference, the AI generates or modifies the image of the garment to realistically fit the given pose, ensuring that the garment maintains its natural shape and drape. 

Benefits and Applications 

The use of ControlNet has brought significant benefits to our image creation and visualization process: 

• Consistency in visualization: We ensure that garments are represented consistently across different poses and bodies, improving the reliability of visualizations for end users. 
• Enhanced realism: By accurately maintaining the shape and pose, the generated images offer a level of realism that enhances the customer experience, facilitating the purchase decision. 

 

Conclusion 

At hiberus, we are just beginning to explore the possibilities offered by generative AI in the realm of fashion. Each project we work on not only pushes us forward in terms of technological innovation but also redefines what is possible in terms of customer experience. We are excited about what the future holds and committed to continuing to lead the way in applying generative AI to revolutionize industries. 

 

Do you want to harness the power of Generative AI to boost your business? We have a team of experts in Generative AI and Data who have developed GenIA Ecosystem, an ecosystem of proprietary conversational AI, content and data generation solutions adapted to the needs of each market and client. Contact us and we will be happy to help you.