Welcome to AI Agents for Image and Video Generation, built in partnership with Google. Text is only one way humans communicate. Most of what we experience is through images, video, and audio. In this course, you'll build agents that produce images and videos for real applications like building website assets and product demos. The user would describe what they want, provide a style guide, and the agent generates options, evaluates them and iterates until the output passes a quality threshold. Generating an image from a prompt can seem straightforward, but building a system that consistently produces high-quality results at scale with many different images and perhaps videos is a different challenge entirely. The main bottleneck is usually evaluation or evals. How do you know if an image matches the brand guidelines and how do you verify that the video maintains visual consistency across multiple scenes? This course teaches you different evals for building image and video agents. I'm delighted to introduce the instructors for this course, Katie Nguyen and Wafae Bakkali. Katie is a developer relations engineer at Google Cloud AI and Wafae is a generative AI technical lead at Google. Thanks Andrew. You will learn three evaluation techniques. First, you will use image text similarity metrics like SigLIP. SigLIP is a dual encoder embedding model that independently encodes text prompts and images into a shared embedding space. Then it computes similarity scores between zero and one. You will use this to see if the output image is aligned with the input prompt. Then you will use LLM-based judges to define custom criteria in natural language and get scores, explanations, and actionable feedback that can be fed into the next generation of the agent. This approach excels at questions that can't be encoded as fixed metrics. Finally, you will implement structured rubrics. The agent decomposes a prompt like, a cat sitting on a windowsill into specific verifiable questions. For instance, is there a cat in the frame? And is the cat sitting? Then it scores every question to measure how well the outputs align with the original prompt. This way, you know exactly which elements passed or failed. By the end, you'll build two complete agents. The first generates UI mockups from brand guidelines. It analyzes style guides, creates design concepts, generates images, and evaluates them against brand consistency. The second produces multi-scene explainer videos. It plans scenes, generates reference frames, animates them, evaluates temporal consistency, and concatenates them into a finished video. In the last lesson, we will put all of the concepts from the previous lessons into different agent skills files and build generative media agents using Gemini CLI. Many people have worked to create this course. I'd like to thank Holt Skinner from Google Cloud for his contributions, and from DeepLearning.AI, Esmaeil Gargari and GT Wrobel. The next lesson covers what generative media is, the model behind it, and the architectural paradigms that power these systems. At the end of the lesson, you will get a clear mental model of the generative media landscape to help you understand which models to use for different tasks. That sounds great. Let's go on to the next video and get started.