Skip to Main Content

Research Guides

Artificial Intelligence for Image Research

A guide on how to use Generative AI for image generation, editing, concept creation and development.

Generative AI Image tools sorted by uses:

There are a ton of image-generation AI tools that span a variety of uses. These tools have different strengths, biases, and features that make them suitable for different use cases.

Note: AI tools and their pricing models may change from the time this guide is posted

Concept Exploration:

Style favors art and illustration. Requires an account and the Discord app. Requires payment.

Style favors photorealism, and features in-painting (editing selected image areas via prompts). Requires payment.

Final Visualizations:

A free and open-source model that is highly customizable. Can be run through multiple different platforms, even your own computer if the hardware supports it.

  • Stable Diffusion on Replicate - Accessing Stable Diffusion through the Replicate UI
  • Stable Diffusion on Automatic1111 - Accessing Stable Diffusion through a downloaded UI, allows for further customization
  • ControlNet - ControlNet is a neural network structure to control diffusion models by adding extra conditions, allowing for precise control over image generation such as specifying poses, objects, lines, etc. via user-drawn inputs.

Architecture-specific AI model that can create render images with user-specified massings. Able to create either rendered visuals or model photos using a guide image to define the building shape. Includes a Rhino plugin. Free credits are issued daily.

Style Transfer:

Free trial. Includes a variety of different Machine Learning models, such as style transfer, super-resolution, video creation, music creation, etc.

Collection of AI models such as text-to-video, text-to-image, and image-to-image, with the ability to train your own models. Free with basic features.

Concept Training:

AI platform that allows for you to train concepts (see: LORA - Low-Rank Adaptation), that act as guides for image generation. Useful for guided image generation based on your dataset of imagery.


3D mesh generator that creates mesh models from text prompts, importable as multiple file formats. Free.

3D mesh generator that creates mesh models from text prompts, importable as multiple file formats. Able to specify styles (realistic/low-poly). Free with limited credits on sign-up.

Image Editing:

Enables the ability to add, remove, or expand content in images through text prompts in Photoshop. Trained in Adobe Stock photos. Requires paid Adobe subscription.


Free image upscaling, is useful to sharpen edges and clarify unresolved details. Free allows for 2x upscaling, paid goes up to 8x.

Paid image upscaling running on replicate, can upscale from 2x to 10x the original image resolution.

General Purpose/Other:

Free and easily accessible, supersedes Dall-E Mini.

Allows users to visualize the patterns learned by a neural network (free with limited credits)

AI model specializing in text generation, an area AI image generators typically struggle with. Free to use.

An open-source AI playground where users can do direct comparisons of a prompt across different AI models


For advanced users:

Generative Adversarial Networks (GANs)

a type of conditional GAN that is trained on paired image data, meaning that it requires both input and output images during training. It is commonly used for tasks such as image segmentation, where the goal is to convert an input image into a corresponding output image with specific labels. For example, given an image of a street scene, Pix2Pix can be trained to generate a corresponding image where each pixel is labeled as "road," "car," "tree," etc.

an unpaired image-to-image translation model that can learn to convert images from one domain to another without explicit pairings. CycleGAN is often used for tasks such as style transfer, where the goal is to convert images from one artistic style to another (e.g. turning a photograph into a painting).

Additional introductory reading on Pix2Pix and CycleGAN