In the following article, I will refer to Midjourney as MJ and Dall-E 2 as Dalle2, at first I wanted to write a solid, long comparison, but then I’ve decided to stick with a quick but straightforward review based on my personal experience.
I've heard a lot of cool things and have seen a lot of breathtaking visuals generated by MJ, but to be honest, I was too busy and invested in Dalle2 (and still am) to spare some time to try a totally different product. What made me intrigued was just another Instagram post referring to Dalle2 as "dull" compared to MJ. The next thing I knew, I was already reading their documentation. It took me a little while to properly follow their discord channel and prompt-generating ways, but Dalle2 is pretty straightforward and easy to use for me.
Pricing and subscription
The recent announcement from OpenAI left many people unsatisfied, the main issue being that you have to pay for generations that may or may not be chosen by the user - so you still have to pay $0.032 per image. It kind of feels like a gamble from the user's side, since you don't know four images how many will answer your prompt perfectly. OpenAI later shared that they are open to suggestions and are considering some changes in the future.
— “Artists who are in need of financial assistance will be able to apply for subsidized access. Please fill out this interest form if you’d like to be notified once more details are available.
As for MJ, if I had to choose someone who uses it rarely as a hobby, I'd choose their Basic plan (you'll be able to run around 200 jobs per month). Upgrading your plan is available anytime, and you will receive a discount proportional to the GPU minutes you haven't used. For example, you'll receive a $5 discount if you use 100 out of your 200 minutes. MJ's pricing plans are more diverse compared to Dalle2, but there's still a good chance for it to change since the current version is in Beta.
above: Dalle2 update
below: MJ pricing
First Impressions
Not going to lie, I was in awe when I saw what MJ could do. I can definitely see why so many people find it to be a powerful tool for creating mesmerizing art. Now, enough about first impressions; let's explore them in depth.
3 prompts, 3 notices, 3 SIDE-TO-SIDE comparisons…
In this part, I will compare these two by giving them the same prompt and evaluate how accurately they follow the prompt and how many generations will be needed to get the final result.
Prompt 1: A ukiyo-e style drawing of a scenic life from a Japanese market.
left: MJ right: Dalle2
Indeed MJ’s variant is more aesthetically pleasing than the one generated by Dalle2 but let's check how we came to this. MJ took 2 parent image threads to get to this result.
MJ, from left: Parent, child, child
dalle2
I will evaluate them based on style and prompt accuracy, in terms of style, MJ was way better than Dalle2, but I’d say Dalle2 followed my prompt more accurately, if I added medium to Dalle2’s prompt, something like “woodblock prints, in style of Hokusai, vintage” it would generate better version on the very first generation
giving Dalle2 more specific prompt
I have this feeling that MJ is way better when it comes to giving vague prompts (randomization works better with MJ) and then upscale until you feel satisfied, while Dalle2 is like a hard-working student, that does its job right but lacks in aesthetics.
Notice 1: with MJ, after a couple of variations the overall content quality drops, just like in the case of Dalle2’s variations.
Prompt 2: Scenes of river in the foreground, sun setting in the background, Sigma 14mm f/1.8 DG HSM Art.
I wanted to check the water reflection accuracy by MJ, since I’m super satisfied with how well it is executed by Dalle2. (trick: instead of using adjectives like “hyper-realistic, super realistic, photorealism, 8k and etc, use certain lenses to get better quality, check my article on this matter)
left: MJ, right: Dalle2
Here is the parent thread, you can see that MJ does these weird tweaks and I’m sure if I’d given adjectives instead of the lenses, it would provide a better result. While all four images generated by Dalle2 executed the prompt pretty well.
Dalle2 generations
Notice 2: On the first generation, all four images of Dalle2 tend to be different from each other (in terms of style, colors, perspective, medium etc), while MJ generations tend to be similar to each other.
Prompt 3: A huge insect which has a very detailed face of a sad man and is wearing a salesperson uniform, medium: black pen, vivid yellow background, Zdzisław Beksiński
left: MJ, right: Dalle2
As you’ve guessed, my inspiration for this one was Gregor Samsa, god bless.. I’ve recently published a small project, where I redesigned covers for my 9 all-time favorite books.
Notice 3: MJ got Zdzisław Beksiński’s style perfectly, but lacked in terms of giving a variety of content, the original thread felt like variations of each other, rather than 4 different images. Dalle2 has been more versatile but wasn’t very close to Beksiński’s style.
Conclusion
It won’t be fair from my side to just say which one is better since I haven’t spent much time testing MJ but what can I say, I was shocked by the aesthetics MJ provides, in their tips for text prompts we read: “but anything you leave out will be randomized. Being vague is a great way to get variety, but you may not get what you’re looking for.” Honestly, this is the best way to describe my first try with MJ. If I had some time I’d love to play around with MJ more, but I don’t see myself being an active long-term subscriber of it, when it comes to Dalle2 it may lack in terms of providing aesthetically pleasing shots, but with correct prompt, it always delivers high quality both in terms of style and prompt accuracy.
Back to Top