Guides: AI Images: Creation and Analysis: Image Creation

Image generators

There are different tools to create AI images. Some of the best known are:

ChatGPT
Uses OpenAI's DALL·E for image generation, trained on a large dataset of publicly available text-image pairs collected from the internet. These datasets include a wide variety of visual and textual content to teach the model how to associate text descriptions with corresponding visual elements. DALL·E generates images based on natural language prompts, creating synthetic images rather than drawing from specific databases. It includes safeguards to avoid reproducing copyrighted works or materials created after 1912. The model is designed to handle a range of prompts, though it may encounter difficulties with overly complex or ambiguous ones. Ethical considerations are incorporated into its design to respect intellectual property and promote responsible use.
Microsoft Copilot
Uses Stable Diffusion, a model designed to generate images across a wide range of styles and subjects. It uses the LAION-5B dataset, which contains billions of image-text pairs, to ensure a broad and diverse understanding of visual elements. Stable Diffusion incorporates several safeguards to ensure responsible and ethical use. These include content moderation filters to block explicit or harmful material, rigorous testing by "red teams" to identify and mitigate risks, and adherence to community guidelines promoting ethical use. The datasets used are curated to minimize harmful or biased content, and the model undergoes regular updates to enhance safety features. These measures collectively help prevent the generation of inappropriate content and ensure the model is used responsibly.
Krea.ai
Uses Flux.1, anewer model that excels in photorealism and high-quality general-purpose image generation. It uses a latent diffusion model and has a massive training dataset
Midjourney
Paid subscription only. This model focuses on artistic and stylized image generation, often used for creative projects.

Writing a good image prompt

Be specific and detailed
- Don't say "a dog", say "a fluffy large black and white dog"
Use at least six descriptive words (prepositions and articles don't count)
Include as many of the following components as possible: image type, subject, context, details, lighting, framing, lens and camera, and style
- Image type: Is it a portrait, a comic, a drawing, a photograph, 3D render, etc.?
- Subject: Who/what is in the image?
- Context: Where does the image take place?
- Details or actions: What are they doing? What's going on?
- Lighting: Lighting helps set the mood. Is the scene taking place at night, during the day, inside, or outside? If inside, is there a lamp, a window, neon light, etc.?
- Framing (Perspective and Composition): Is it a close-up, a wide shot, a bird’s-eye view, or a specific angle?
- Lens and camera: If asking for a photograph, is it a wide or tight shot? What kind of lens are you looking for?
- Style: Are you looking for an image inspired by a painter? Or maybe manga, or futuristic, cyberpunk, art deco, etc.?

Create a portrait of a baby girl sitting in a highchair in front of a birthday cake, natural lighting, in the style of Degas (painter).

Image created with MS Copilot.

Prompt engineering is a iterative approach. If you are not happy with your image, rework your prompt and rerun it.
Possible prompt starts as suggested by Microsoft Copilot:

Create an image of…
Generate a vast landscape of…
Create a lifelike portrait of…
Generate a comic book scene with…
Draw an abstract piece of art of…
Generate an artwork of…
Illustrate a (subject or scene) with…

Envision a futuristic design of…
Craft a vibrant panorama where…
Create a painting of…
Create an imaginative scene of…
Conceptualize a cutting-edge design of…
Generate an eco-conscious image of…
Show me an up-close image of…

Why are AI image generators so bad at hands and faces?

Training Data Limitations: AI models are trained on large datasets of images, but these datasets may not have enough high-quality examples of hands and faces in various positions and lighting conditions. This lack of diverse and detailed examples can lead to inaccuracies.
Small Size in Images: In many images, hands are relatively small compared to other parts of the body. This makes it harder for the AI to learn the fine details needed to generate realistic hands.
Overlapping and Occlusion: Hands often overlap with other objects or parts of the body, creating occlusions that make it challenging for AI to understand and generate them correctly.
High Detail Density: Both hands and faces have a high density of details, such as facial features and finger joints, which require precise rendering. AI models sometimes struggle to capture these details accurately.

Image created with Canva. The red circles illustrate mistakes in the hands and faces of the people in the image.