Comparing AI image-generators: Stable Diffusion, OpenAI DALL-E-2 and MidJourney

Recently I’ve tested three popular image generators in order to find our their limitation and see which one is the best on the market at the time.

what are AI image models? AI image models are algorithms that use artificial intelligence (AI) to analyze and manipulate digital images. These models use machine learning techniques, such as deep learning neural networks, to recognize patterns and features in images and make predictions or classifications based on those patterns. (answered by chat.openai.com)

The first one to try was Stable Diffusion. Available as an Open Sourced AI model at the link: https://huggingface.co/spaces/stabilityai/stable-diffusion

Stable Diffusion allows to construct photography-based collages. At the time resulting photo-images looks somehow unnatural, ugly and by many times are missing or adding redundant details.

Stable Diffusion, image constructed from description “white rabbit running in front of Eiffel tower”.

However, Stable Diffusion engine manages well with producing oil-painted or naive art.

Stable Diffusion, image from description “portrait of Picasso”

The next one tro try was DALL-E-2 by OpenAI. This is an evolution of DALL-E model, constructed by OpenAI. Their textual language processing models are hitting today as well as speech recognition models. Surprisingly DALL-E-2 didn’t followed the quality of the other AI models of that company.

DALL-E-2, description “portrait of Picasso, realistic, photography quality details, in colour”

DALL-E-2 seemed to be a step forward comparing to Stable Diffusion. However it’s application is still seemed to be limited to oil-painted like drawings and naive art.

DALL-E-2, image from description “portrait of Saint Ekzuperi while he is in a plane cockpit, sketch style, black and white, monochrome”

The third one engine called MidJourney outrulled the previous two models in times. It looks like it is not aimed to producing collage photographies as well, but instead centres on 3D or realistic art images. However results were quite good for oil-paint and sketch styled images.

MidJourney, image from description “portrait of Saint Ekzuperi while he is in a plane cockpit, sketch style, black and white, monochrome”

MidJourney, image from description “portrait of Picasso, realistic, photography quality details, in colour”

The resulting quality of MidJourney generator is seemed to be much better comparing to Stable Diffusion and DALL-E-2.