Generative AI Empowers Fotor to Create Powerful Image Editing and Design Tools, Elevating Both Efficiency and Creativity

For over 15 years, Fotor has been at the forefront of image processing technology, consistently driving product innovation and enhancing user experience to empower everyone to create studio-quality content more efficiently. With the advent of generative AI, the image editing industry has entered a new era of opportunities. Fotor has taken the lead in integrating generative AI technology into image editing and design workflows, successfully implementing various applications including text-to-image, image-to-image, text-to-video, image-to-video generation, and video editing capabilities. Users can leverage Fotor's comprehensive suite of editing tools for background removal, special effects, and design modifications on their image assets. The output can be seamlessly applied across various commercial scenarios such as e-commerce advertising, social media marketing, and PowerPoint presentation generation.

However, while generative AI has enhanced Fotor's competitive advantages in terms of product functionality, business models, and cost-effectiveness, it has also presented several challenges:

• How to reduce inference latency and enhance the processing efficiency for high concurrency demands in scenarios involving 600 million global users with round-the-clock high concurrency needs;
• How to minimize manual effort while economically, efficiently, and intelligently tagging image and graphic assets;
• How to conduct comprehensive and intelligent reviews of user-generated content to filter out non-compliant or inappropriate content, and respond quickly to any issues;
• Additionally, integrating generative AI applications into existing business systems and developing new models present challenges to Fotor's current capabilities.

Fotor's partnership with AWS dates back to 2014, and since then, all of their critical backend applications have been deployed on AWS. In the era of generative AI, Fotor aims to leverage AWS's comprehensive generative AI capabilities and expertise to empower their business operations, enhance team efficiency, and address the Company’s challenges, all while building upon AWS's robust infrastructure.

As a one-stop image editing and design platform serving 600 million users globally, Fotor's services are deployed across multiple AWS regions including US East, US West, Jakarta, Frankfurt, and São Paulo. The platform experiences relatively stable traffic patterns, with demand fluctuations between peak and off-peak periods not exceeding 25%, while maintaining consistently high concurrent user loads. One of Fotor's key challenges was addressing inference latency in high-concurrency scenarios. By implementing Amazon SageMaker's asynchronous inference capability, integrated with Amazon SQS, Amazon SNS, and Amazon EC2's scheduling capabilities, Fotor efficiently handles high concurrent user demands through asynchronous inference processing. This solution has resulted in more than 50% faster inference times, reducing processing time per request from 10-20 seconds to 7-8 seconds and significantly decreasing user wait times.

Image annotation is a crucial business scenario for Fotor. Previously, this task relied heavily on large-scale outsourced manual labor, where workers would tag images across multiple dimensions including characteristics, colors, and aesthetic elements. However, this manual approach faced several challenges including worker fatigue, declining efficiency over extended periods, and excessive reliance on predefined templates, leading to suboptimal annotation results. Now, for Fotor's international operations, they leverage LLMs via Amazon Bedrock to annotate images and visual content. This AI-powered approach not only significantly improves annotation efficiency compared to traditional manual methods, but also enhances the accuracy and richness of the tags. This advancement has greatly improved the generalization capability of models trained on these annotated materials, resulting in generated content that is both more diverse and accurate.

When users utilize LLMs for text-to-image or text-to-video generation, Fotor can intelligently handle cases of incomplete inputs or prompts. By analyzing users' historical data and current inputs, Fotor discerns user preferences and intentions and employs LLMs combined with prompt engineering techniques to execute semantic expansion. This optimization process helps generate images or videos that precisely align with users' requirements and creative intent.

In image-to-image scenarios, Fotor leverages LLMs to reverse-engineer the original image and extract semantic information that users haven't explicitly described. This inferred information then guides further editing, allowing the generation of images that more closely meet user expectations through semantic expansion based on the original image.

Efficient review and filtering of LLM-generated content remains a pervasive challenge in generative AI applications. When users create images using Fotor, there exists the risk of generating illegal or inappropriate content. Therefore, Fotor requires a robust content moderation mechanism for user-generated images to filter out content that doesn't meet compliance requirements. Leveraging Amazon Rekognition, a cloud-based AI service, Fotor can automatically perform image recognition and video analysis cost-effectively. The service uses pre-trained and customizable computer vision capabilities to detect and analyze over 10 categories of prohibited content, including violence and explicit material. Amazon Rekognition extracts information and insights from images to provide specific confidence scores for moderated content. For instance, if an image is flagged for violent content, the service might indicate a 96% confidence level in that assessment. This enables Fotor to swiftly make content moderation decisions without developing proprietary AI models, thereby streamlining the review process by referencing these confidence scores.

Fine-tuning based on open-source models is one of the ways Fotor applies LLM technology, but it poses two key challenges. Firstly, open-source models often have unclear intellectual property rights. Secondly, initial tests reveal that, even after multiple optimizations, the content generated by these models lacks the image-text accuracy and aesthetic quality needed to meet Fotor's business standards. To address these issues, Fotor utilizes Amazon SageMaker's model training capabilities to integrate their deep expertise in image editing and extensive data resources to expedite the development of proprietary image and video models on open-source foundations. This approach enables them to craft custom solutions with defined intellectual property rights while building upon existing open-source frameworks.

Since the introduction of generative AI in October 2022, Fotor has steadily rolled out hundreds of new generative AI-powered features. These include diverse templates for text-to-image generation, preset parameter templates for image-to-image and text-to-video conversions, and customized workflows tailored to specific user scenarios. With these new capabilities, users no longer need complex image editing software; instead, they can leverage Fotor's comprehensive generative AI functions and convenient templates, thus independently creating desired images and videos at low cost. For international operations, Fotor has enriched its creative features by integrating LLMs via Amazon Bedrock and optimized its existing image editing and design workflows. This improvement has resulted in a 20% increase in satisfaction among international users.

With 600 million users worldwide, Fotor encounters constant high concurrent demand, where any processing delays could adversely impact user experience. After implementing Amazon SageMaker's asynchronous inference capability, Fotor can now handle up to 300 concurrent requests per second, substantially reducing user wait times and enhancing the user experience.

By leveraging Amazon Bedrock to integrate Large Language Models (LLMs), Fotor has successfully introduced innovative creative features that significantly enhanced the international user experience. This technological advancement has driven Fotor's business expansion, resulting in a tenfold increase in daily active users and substantial revenue growth.

With the advent of generative AI, human resources previously dedicated to repetitive basic tasks, such as content tagging and moderation, can now be freed up. These resources can be reallocated to more creative positions, enabling organizations to achieve multiple-fold improvements in overall workforce productivity.

Fotor's collaboration with Amazon Web Services on generative AI is still ongoing. In addition to utilizing Amazon SageMaker's training capabilities during their custom model development process, Fotor plans to implement additional Amazon SageMaker features starting in the future, including data cleansing, training, and deployment functionalities. After implementing these features, Fotor expects to reduce its experimental cycle time from one month to a week, significantly enhancing the efficiency of its proprietary model development.

Founded in 2009, Fotor has been deeply engaged in the image-processing industry for over 15 years. As of July 2024, the platform has amassed a global user base of 600 million users. Fotor is dedicated to providing creative image editing and design tools and services to users across more than 200 countries and regions, including the United States, Europe, Southeast Asia, and India.