123RF Achieves 95% Cost Reduction on Content Translation and Up to 100x Faster Content Screening with Generative AI on AWS

Established in 2005, 123RF offers a vast collection of royalty-free stock images, both generated in-house and sourced from external contributors. Recognizing the shift towards digital creativity, the company has since embraced artificial intelligence (AI), expanding its services to include a range of innovative AI tools like text-to-image generators, background blurring, and 123RF AI Writer to meet the evolving demands of the creative industry. Novianto Kho, senior IT manager at 123RF, says, “We’re transforming our company into an AI-based enterprise, with the vision to empower creators with real freedom.”

Since its inception, 123RF has been committed to ensuring its images are appropriate and comply with copyright regulations. The business chose to host its operations on Amazon Web Services (AWS) from the start. Initially, the company relied on a team of over 20 staff members to engage in careful content moderation. Subsequently, 123RF developed machine learning (ML) models using open-source technology to automate image screening, which are hosted on Amazon Elastic Compute Cloud (Amazon EC2).

However, the models and output did not detect all content violations, which resulted in the company assigning more employees to screen images. To fully automate content moderation and save time and resources, the company consulted with AWS. Additionally, it sought to improve the accuracy of ML models in flagging unsuitable content.

123RF subscribes to AWS Enterprise Support and holds weekly meetings with its technical account manager to collaborate on technology solutions that advance the business. "AWS recommended we test Amazon Bedrock, a fully managed service for building generative AI applications. In just three weeks, we completed a successful proof of concept (POC) and implemented Anthropic Claude 3 models on Amazon Bedrock," states Kho.

123RF engineers first configured their virtual machines, then assigned appropriate security roles, and finally installed the software development kit for Anthropic Claude 3. Kho states, “Implementing Anthropic Claude 3 was quick and straightforward, and we utilized AWS expertise to expedite the process.”

Prior to implementation, 123RF worked with AWS to test a subset of several million pieces of user-generated data, to verify the accuracy of Anthropic Claude 3 models in flagging unsuitable content. Though 123RF quickly determined that Claude 3 fits its use case, Kho recommends that companies experiment with different pre-trained models and ML tools on AWS to customize a solution per use case. Kho also emphasizes the importance of continuous learning. “Because AI technology—including generative AI—evolves so rapidly, investing in skill development is essential,” Kho says. In addition to working closely with AWS solutions architects, 123RF employees have attended AWS classroom training sessions on specific ML topics.  

Since implementing Anthropic Claude 3 for content moderation, 123RF has explored additional use cases, including content translation. Although the company achieved significant success in English-speaking markets with its generative AI–based semantic search tool, it faced content discovery challenges in 15 other languages due to English-only titles and keywords. The cost of using Google Translate for continuous translations was prohibitive, and other models like Claude Sonnet and Open AI GPT-4o were not cost-effective. Although Open AI GPT-3.5 met cost criteria, it struggled with consistent output quality. This prompted the search for a more reliable and affordable solution to enhance multilingual content discovery.

Initially skeptical due to previous cost optimization challenges, the 123RF team, together with AWS solutions architects, conducted rapid iterations and small POCs on Anthropic Claude 3 to refine translation prompts. 123RF successfully implemented the solution into production in July 2024 and has since reduced batch translation costs from $120,000 to $6,000 for nine languages.