In the rapidly evolving world of web graphics and social media, Artificial Intelligence (AI) image generators have emerged as a promising alternative to traditional stock photography. With a plethora of tools available, we embarked on a journey to compare the capabilities of six distinct AI text-to-image generators. Our findings shed light on the differentiators and considerations to keep in mind when harnessing these innovative tools for creative endeavors.
The Landscape of AI Image Generators:
The current market is flooded with a multitude of AI image generator options, with new contenders entering the scene regularly. For this study, we selected six prominent AI tools, acknowledging that there are many more available. Each of the chosen tools possesses the remarkable ability to translate textual descriptions into captivating visual representations. However, it’s crucial to recognize that among these tools, certain differentiating factors stand out, and none are as pivotal as the terms of service they offer.
Navigating the Terms of Service
It’s essential to note that every AI image generator comes with its unique set of terms of service. Ranging from permissive to restrictive, these terms can significantly impact the practicality of using these tools. Restrictions may include mandates for attribution or the disclosure that the image was AI-generated and not real. Some tools may necessitate a subscription for larger organizations, others are freely accessible, and one even requested a donation in a cryptocurrency. Since terms of service are susceptible to change, we recommend meticulously reviewing them prior to investing substantial time in learning and utilizing an image generator.
Review Methodology
Our evaluation centered on the depiction of a tyrannosaurus rex, a task that highlighted unique challenges for each generator. Specifically, we assessed their ability to portray the T-Rex’s distinctive short forearms and intricate finger (claw) detailing. We presented the complex scenario of a T-Rex donning dark sunglasses and a white lab coat, situated within a biology lab while holding a petri dish. Notably, some inmage generators struggled with the sunglasses’ shading, while others exhibited uncertainty about the appearance of a petri dish.
The Vision: A Happy T-Rex in a Pixar-Style Render
Transforming the fierce tyrant lizard into a joyful character was an intriguing experiment. Our request aimed to render the T-Rex in a Pixar-style image, exuding happiness. This imaginative challenge prompted us to consider how a historically fierce creature could be reimagined in a cheerful light.
Our evaluation adhered to the default settings and presets of each tool. While some platforms offer extensive options for resolution, aspect ratio, variability, and more, we maintained uniformity by refraining from adjusting these settings. It’s important to recognize that even minor changes in such parameters could yield drastically diverse outcomes.
In alphabetical order:
Adobe Firefly (Beta)
The sunglasses and lab coat are correct, and the arms are shorter than most of the generators tested. However, Firefly seemed to struggle with the fingers (claws) and didn’t get the petri dish correct.
It is important to note this tool is in beta, and has probably the most restrictive terms of use we’ve seen. When Firefly is out of beta, we hope that changes.
tyrannosaurus rex, dark sunglasses, biology lab, white lab coat, holding petri dish, pixar style, happy mood
Bing
Bing was the surprise of the bunch. It is based on DALL-E-2, yet the output looked completely different.
In two of the images it generated the sunglasses correctly by adding shaded lenses instead of clear frames. In the two other images with clear glasses, it generated an image closer to a petri dish, but it’s not quite there yet. The claws seem to be the most anatomically correct, though the arms seem to be a bit big for a T-Rex.
tyrannosaurus rex, dark sunglasses, biology lab, white lab coat, holding petri dish, pixar style, happy mood
Canva
https://canva.com/ai-image-generator
Canva’s generate produced some of the lease accurate results. One of the images does not feature a T-Rex at all, and the image generator did not include accurate depictions of petri dishes.
As Canva learns with more images it can only improve, but their competitors definitely have a head start.
tyrannosaurus rex, dark sunglasses, biology lab, white lab coat, holding petri dish, pixar style, happy mood
DALL E 2
We featured a blog post on DALL E 2 in February, but this is the first time we asked for a cartoonish image (Pixar style).
This was the only generator to correctly put dark sunglasses on all four images and also depict a happier feel. Note that in some cases it had problems with claws and petri dishes.
It is interesting to note that the Bing generator is based on DALL E 2 despite being given the same prompt, the results were quite different.
tyrannosaurus rex, dark sunglasses, biology lab, white lab coat, holding petri dish, pixar style, happy mood
Midjourney
This generator best understood the “Pixar style” of animation. It also rendered the fingers (claws) pretty well, though the arms are too large for a T-Rex. Half of the images are in a biology lab setting.
Unfortunately Midjourney struggled with the dark sunglasses and the petri dish.
tyrannosaurus rex, dark sunglasses, biology lab, white lab coat, holding petri dish, pixar style, happy mood
Stable Diffusion
https://StableDiffusionWeb.com
This generator missed the mark almost completely. The prompt had to be modified slightly (note the insertion of “cartoon”), and still this was the best output. None of the images feature a T-Rex as we requested, but at least they are all in a biology lab setting.
There is a newer version, Stable Diffusion XL, but the server was too busy. Our request was waiting in the queue all day.
Published samples on their web site look great, but the output our tests generated never got close.
tyrannosaurus rex, dark sunglasses, biology lab, white lab coat, holding petri dish, cartoon, pixar style, happy mood
Giving it a try? A few tips before you start:
- Every generator tested will create four images. If you like any of the images, grab it immediately. The same prompt on the same AI generator won’t produce the same result a second time.
- If one of the four images is close to what you like, you can usually generate variations on that one image. A variation is like saying, “more like this one.” How this works varies from tool to tool.
- Many tools have a library of images that were generated by others, along with the prompts that created them. This is an excellent way to learn the specifics of the prompt for that generator without exhausting your allotment of images. It can be quite entertaining to see what others have created!
- Some generators have built-in guard rails to keep trademarked images and likeness of celebrities out of the generated images, some do not. Be sure you read the terms and know what the liabilities in using the images.