Driven by urgent need to streamline the banner image generation flow, the company selected Adimen to implement its hands-on expertise in Al-backed automation. To fulfill the project’s objectives we provided Al algorithms based on Machine Learning trained models that could effectively resize the original templates into a number of required sizes.
Our solution enabled the following functional advantages:
1. The original form factor was built manually in Google Web Designer with the resolution of 300×250 and set to be processed for further resizing.
2. Al techniques like image recognition and object detection were utilized to sort out all the template el-ements.
3. Background as well as other image objects were thoroughly processed to not factor out some key elements like faces, brand attributes, corporate logos, etc. during resizing. The original template got resized into 728×90, 160×600, 300×600, 120×600 and 970×250 resolution formats.
4. Two additional Facebook-specific image sizes of 1200×628 and 600×600 were generated to serve the social media purposes.
5. Respective fonts and animations were also automatically rendered in accordance with the size of the banner image they belonged to.
Resizing has always been a very tedious and time-consuming process for designers. Along with this, it’s cost a pretty penny for business owners as they’ve had to pay more for extra hours. Moreover, the human factor usually contributes to numerous mistakes and inconsistencies. Earlier on, all this offered little room for optimism.
Implementing our Al-fueled solution provided the customer with the following procedural benefits:
A robust trainable Al model that allows smooth automated resizing of the manually built image into a set of other size formats. The model can be easily retrained if there is a case.
The rendering of the original banner image into another, even bigger size does not imply the loss of quality.
Output animated multi-frame HTMLS banner creatives to be run on desktop and mobile devices inherit all relevant properties of their original templates.
Al algorithms deal directly with datasets stored in a DAM system, thus freeing up a lot of time for human specialists for analysis and iteration.
The applied solution enjoys great flexibility and scalability, meaning that the Al model can be easily retrained to create images with another size.
Tech stack used in the project.
To achieve the best results, the technology stack comprised a number of the most efficient programming tools. The following languages, containers and frameworks were utilized: Python, TensorFlow, OpenCV, Numpy, Flask, YOLO and Docker