Generative AI is a subset of artificial intelligence that focuses on creating new content, such as text, images, music, and code. It is trained on large datasets of existing content and then learns to generate new content that is similar to the training data.

Using deep learning and neural networks, generative AI models learn from vast datasets, enabling them to generate content from their training data. They can perform a wide array of tasks, from crafting blog posts and translating languages to producing stunning visuals and composing melodious tunes.

Text Generative AI

Generative AI models can generate text, such as blog posts, articles, and even novels. They can also be used to translate languages and write different kinds of creative content. Example tools included Chat-GPT, Bard, and Copy.Ai

Visual Generative AI 

Generative AI models can generate visuals, such as realistic photos of faces, objects, and landscapes. They can also be used to create artistic images, designs and videos. Stable Diffusion, Midjourney, and CogVideo for example.

Music Generative AI

Generative AI models, such as Jukebox by OpenAI, Musenet, and AIVA, can generate music, such as melodies, harmonies, and rhythms. They can also be used to create new musical genres and styles.

Code Generative AI

Chat-GPT, Copilot, and Codex are examples of Generative AI models which can generate, debug and optimize code, such as Python, Java, and C++.

As there are numerous generative AI to cover, we will be focusing on Image generative AI in this article. These AI creates of entirely new images based on statistical pixel data. The process involves training these models on extensive image datasets, enabling them to understand and replicate pixel-level patterns.

These models not only generate images from scratch but can also bring textual descriptions to life. They convert text into a latent representation that summarizes essential image features like styles, colors, and frames, using this information to craft images matching the text’s description.

There are 3 main image generative AI models

1. Diffusion Models: These models, also called denoising diffusion probabilistic models (DDPMs), use a two-step process: forward diffusion adds noise to data, and reverse diffusion reconstructs the original samples. Despite longer training times, diffusion models can have numerous layers, leading to high-quality output. They are versatile but slow due to reverse sampling.

2. Variational Autoencoders (VAEs): VAEs consist of encoder and decoder neural networks. The encoder compresses input data, preserving essential information for the decoder to reconstruct it. VAEs generate outputs quickly, although they lack the detail of diffusion models.

3. Generative Adversarial Networks (GANs): GANs involve a generator and a discriminator. The generator creates new examples, and the discriminator learns to distinguish real from generated content. While GANs generate samples quickly, their diversity is limited, making them suitable for specific domains.

As the Diffusion model is one of the most widely-used, we will explain it in more detail. Famous Diffusion model tools are names such as Stable Diffusion and Midjourney. They are both text-to-image diffusion models, but there are some key differences between the two.

Midjourney

• Closed-source proprietary model, developed by Midjourney Inc.
• Only accessible through the Discord chat app
• Easier to use
• Has a more limited set of features
• Costly starting at $10 per month to $120

Stable Diffusion

• Open-source model, anyone can use it
• Can be used on a variety of platforms, including online and offline
• Comparatively a bit more difficult to use at the expense of more customization
• Has a wider range of features, such as image-to-image functionality
• Free!

Overall, Stable Diffusion is a more versatile and powerful model than Midjourney. It is also more accessible and affordable. Not to mention the wide range of available community-trained models to generate unique visual style. For these reasons, Stable Diffusion is generally considered to be the better model of the two.

Real-World Use Cases

Image generative AI has made a significant impact in various real-world scenarios. In the realm of advertising and marketing, it’s employed to create eye-catching visuals that effectively convey the value of products and services. Architects and urban planners use AI-generated images to fashion realistic representations of buildings and urban spaces, aiding in the design and presentation of projects. Some artists are even embracing AI-generated elements in their work, blurring the lines between traditional art and AI-powered art. Education industry can also greatly benefit from AI-generated images, making materials more engaging and visually-appealing. It can be utilized to let the students’ creativity run wild!

Now, you might be wondering, if it is so good, how could I also use it? Well, the good news is that, Stable Diffusion is free!

Here is a brief step-by-step on how to use Stable Diffusion.

How to use Stable Diffusion

1. Install Stable Diffusion. 

You can install Stable Diffusion on a variety of platforms, including Windows, Linux, and macOS. However, for it to be easy to use, a UI is needed, and most of the time the community developers often use Web UI. If you are not very well deep into this topic, I recommend a Stable Diffusion AUTOMATIC 1111 package as it is easy to install and use. There are also several online websites that you can use to run Stable Diffusion without having to install it locally.

2. Download a Stable Diffusion model

There are many different Stable Diffusion models available, each with its own strengths and weaknesses. You can download models from the Hugging Face website or from the official Stable Diffusion GitHub repository.

3. Run Stable Diffusion 

Once you have installed Stable Diffusion and downloaded a model, you can run Stable Diffusion to generate images. To do this, you will need to provide a text prompt. The text prompt should describe the image that you want to generate.

4. Adjust the parameters. 

Stable Diffusion has a number of parameters that you can adjust to control the output of the model. Some of the most important parameters include:

Sampling steps: The number of sampling steps controls how detailed the generated image is. Higher values will result in more detailed images, but they will also take longer to generate.CFG scale:

The CFG scale controls the strength of the guidance from the text prompt. Higher values will result in images that are more faithful to the text prompt, but they may also be less realistic.

5. Generate the image

Once you have adjusted the parameters, you can generate the image by clicking the “Generate” button. Stable Diffusion will generate the image and display it in the window.

For example, if I run Stable Diffusion and provide the following text prompt: “A red panda riding a bicycle down a busy street,” here is how it would look like.

Stable Diffusion is a powerful tool for generating and editing images. It is still under development, but it has already been used to create some impressive results.

Real usage

In a recent university project for my JM421 class, I successfully created nearly 100 captivating visual assets and these images played a vital role in enhancing the overall appeal of my educational media project. It immensely helped me in crafting visually engaging concept art. Its use case was for education.

However, if it is for commercial, it is worth noting that there are important ethical considerations to take into account as the automation of creative processes raises questions about the boundaries of authorship and creativity, and these concerns should not be overlooked. It is crucial to address the ethical implications associated with its use as the AI-generated images copyright and ownership is still in debate. Thus, one should utilize it carefully, and it would be best to hire full-fledged artists until these issues are solved.

–

In the end, Image generative AI models like Stable Diffusion sure have emerged as powerful tools, blurring the lines between human creativity and machine intelligence. Stable Diffusion, with its open-source nature and versatile capabilities, stands out as the latest human innovation, allowing users to seamlessly translate text prompts into any images in any style.

As showcased in this article, Stable Diffusion empowers individuals to explore their creativity, transforming ideas into tangible visual expressions. While the learning curve might be steep initially, the rewards are endless for those willing to learn.

Sure, image generative AI has transformed the way we create and interact with visual content. From concept art to medical imaging, its applications are diverse and far-reaching. However, with great power comes great responsibility. Ethical considerations, including copyright, ownership, bias, and transparency, are still among the top issue regarding AIs.

The world of AI-generated images is still evolving, and as AI models become more sophisticated, their potential for creative and practical applications continues to expand. So, If you feel like learning more about generative AI, there are a wide range of online resource available. The Stable Diffusion Reddit community for example, there are tons of guides available there.

Have fun generating! 🙂

By

Pitchaya Jenjirawong 6307640018
Banthita Boontiam 6307640034
Prawravee Suwantawit 6307640042
Patcharaorn Yokyong 6307640398
Mooktri Kaeng-in 6307640349