Bridging Art and AI: Scaling Photoshop-Based Image Editing to Train Generative AI Models

Image Annotation: Editing 2D Images

The client leveraged our experience in Adobe Photoshop to edit AI-generated images by applying different artistic styles and then modifying specific attributes, such as altering eye color, clothing shades, and other visual elements. This process was designed to diversify the dataset, enhancing its variety for training an image-generative AI. By combining AI’s creative potential with manual adjustments in Photoshop, we were able to produce a wide range of visually distinct images, effectively increasing the dataset’s depth and quality. The client was highly satisfied with the results, noting the improved richness and variability of the dataset, and is planning on further expanding the project further to support future AI training needs.

Companies looking to use AI to create image datasets to train their ML systems usually run into problems when relying solely on AI: the images produced by AI have issues with generating facial features or hands, which will need to be corrected with additional training data. Our client sought to avert these problems and get things right the first time by training their generative AI system with images that do not have these issues and have already been edited in Adobe Photoshop to diversify the training dataset. This included things like eye color, clothing shades, and facial expressions, to ensure a broader range of visual variations. The purpose was to create a large volume of images that would improve the AI’s ability to generate more accurate and realistic visuals. 

In addition to this, they wanted to introduce new generative features, such as incorporating religious attire or conveying different emotions in the images, to enhance the AI’s versatility and output quality. Since this is a very complex and time consuming project, the client was looking for a trusted data annotation partner to help them complete this work so they can focus on their core business needs.

The client chose to work with us because of the strong results our team had delivered in previous Photoshop-driven digital asset (DA) projects. Impressed by our expertise in leveraging Photoshop for complex tasks, the client was interested in exploring whether a team with a deep focus on DA could successfully handle an art-focused project. They saw an opportunity to test our ability to apply our experience in digital asset creation, manipulation, and refinement to more creative, art-oriented tasks. By tapping into our Photoshop proficiency, the client aimed to push the boundaries of their generative AI project while benefiting from our established track record of delivering high-quality, precise results in previous DA work.

Mindy Support successfully sourced and recruited a team of 10 skilled professionals, including data annotators and designers, to support the client’s project. The team utilized Adobe Photoshop extensively for the image editing portion of the work, while the curating and image filtering tasks were handled in the client’s proprietary tool. To ensure that all team members were aligned with the client’s specific artistic vision, we used the client’s own data samples as a foundation for training. This included dedicated training sessions focused on mastering the particular style required for the project. The process of onboarding and training the team was critical to ensure consistency across all deliverables.

One of the most challenging aspects of the project was meeting the client’s high standards of “beauty” in the edited images. The subjective nature of aesthetic preferences made it difficult for data annotators, who were not used to such creative tasks, to quickly adapt to the image editing process. We addressed this challenge by holding multiple meetings with the client to refine a style guide (GL) and conducting additional training sessions. These sessions included hands-on work with sample images to ensure that all team members could produce edits that aligned with the client’s expectations. The result was a seamless integration of both technical and creative efforts, and the client was highly satisfied with the outcome. Given the success of this project, they are planning to expand the current 2D design team with additional 2D artists, growing the team to up to 15 people, and are also testing a new editing approach with 3D designers to further enhance their capabilities.

• Sourced and recruited a team of 10 designers and data annotators
• Quality score 98%+
• 100,000+ images annotated