Adobe released the latest version of its “ethical” artificial intelligence image model this week and with it the promise of greater photorealism and improved prompt following. Midjourney has also seen continual improvements since the release of version six at the end of last year.
Midjourney and Adobe Firefly are among the most widely used AI image generators for different reasons, with Adobe finding a home among traditional artists and designers and Midjourney the darling of a new generation of AI artists.
Both models feature in my best AI image generators guide. To see whether Adobe Firefly and its licensed training dataset can match the quality and prompt adherence of Midjourney and its murkier origins I’ve put them head-to-head with 7 prompts.
Why test Adobe and Midjourney?
With Adobe exclusively training its model on the hundreds of millions of images licensed as part of the Adobe Stock library — including some made by Midjourney — I wanted to see if its promise of ethical sourcing plus impressive output could stand up against Midjourney.
Unlike Adobe, Midjourney does not disclose the source of its training data. The startup likely used images scraped from the open web without permission from the creators and as a result is facing lawsuits from artists claiming it stole their work.
In a lawsuit of its own, OpenAI said it is impossible to train a leading AI model without copyright material in the training data. For now the question of training data, provenance and legality are very much up in the air and open to a multitude of court cases and lawmakers.
Designing the test
This is an interesting way to see if that actually holds true. Can the massive library of anything it could find for Midjourney match the limited but structured and licensed dataset for Adobe?
I created nine prompts that push the AI to adhere to a specific request, are generic enough that it has to work out what to do and specific enough to offer a comparison. All settings were left on default with no specifics beyond making it widescreen.
1. A Quiet Morning in a Venice Canal
Lets start with a stunning scene, something peaceful and historic in equal measure. For this prompt I wanted to see something picturesque but not artistic, however I kept the prompt generic enough not to force a photo or a painting.
The prompt: “Show an early morning scene along a Venice canal with light fog. Small boats are docked along ancient, moss-covered buildings, and a solitary figure in a striped shirt is feeding birds.”
These are two very aesthetically different images. One a simple depiction of a beautiful space, the other an intricate work exploring beauty in decay. Both are compelling images but neither have realistic water. Midjourney wins for quality and detail but I don't dislike Firefly.
2. Cyberpunk Street Market
Lets jump to an alternative universe where cyberpunk is the dominant style. This should create a game-like scene and that is exactly what we got. Lots of neon.
The Prompt: “Visualize a bustling cyberpunk street market at dusk. Neon signs glow above stalls selling futuristic gadgets. Crowds of diverse characters, some human, some robotic, navigate the narrow, rain-slicked streets.”
I was surprised how similar these two images were to one another. The Midjourney image looks like someone took the Adobe Firefly picture and put it through an upscaler, or that it was given longer to render. While I like the mystery of the Adobe image the quality of the Midjourney picture easily wins out.
3. Ancient Library with Hidden Secrets
For this one I wanted to capture the essence of knowledge. A sense of information swirling around the space. I didn't want a picture of a shelf of books.
The Prompt: “Imagine a vast, ancient library with towering wooden shelves filled with dusty books. A grand chandelier hangs above. A few beams of sunlight stream through high windows, illuminating hidden messages in the architecture.”
Both images look like a different perspective on the same room. One a photograph taken by a student their phone, the other by a gifted photographer. The Midjourney image is better, not significantly so, but better and captures the prompt well.
4. Desert Oasis at Sunset
Love Oasis, particularly their earlier material. But for this prompt we're looking to the desert to create a pocket of life in an otherwise desolate space. The idea is that to follow the prompt it needs to capture the concept of abundance amidst isolation.
The Prompt: “Capture a serene desert oasis at sunset. Tall palm trees and lush greenery surround a small, crystal-clear pond. Sand dunes stretch into the distance, glowing orange and pink under the setting sun.”
I didn't like either of these images. Adobe created a cut-scene from GTA-3 and Midjourney made one from GTA-5. Neither truly encapsulated the idea of abundance or isolation but Adobe got closer to the prompt.
5. Victorian Inventor's Workshop
From desolation to imagination. The next prompt is designed to see how well the AI models deal with the concept of order through chaos. I want weird experiments, gadgets and a fun retro vibe.
The Prompt: “Depict a cluttered Victorian inventor's workshop. The room is filled with strange brass machines, gears, and tools. Papers and blueprints are scattered everywhere, and a large, intricate clockwork device dominates the centre.”
Adobe Firefly and Midjourney somewhow both created an image that looked like a photo taken during a museum tour of some old inventors house a century after his death. Midjourney did a better job at captured the full space and whimsy but there is something about the Adobe image that is captivating.
6. Frozen Waterfall in a Magical Forest
Do you want to build a snowman? No, neither do I as I'm not a fan of the snow but there is nothing as magical as a frozen waterfall — from a distance. This prompt is designed to see how well it captures the concept of unreality.
The Prompt: “Picture a large, frozen waterfall cascading through a snowy forest. Icicles hang from the branches of frost-covered trees, and magical creatures peep out from behind the trunks, glowing eyes watching curiously.”
The winner here comes down to personal interpretation. If you like something out of a direct-to-video cartoon horror movie from the 90s then Adobe wins. They both look like something out of a horror movie and neither followed the prompt perfectly.
7. Futuristic City from Above at Night
A future city is almost one of my default prompts to try on any image generator, that and something to do with cats on the moon. Its a good test of prompt adherance as it needs to understand your vision of the future.
The Prompt: “Show an aerial view of a futuristic city at night. Skyscrapers covered in digital screens tower over glowing streets filled with flying cars. Parks and rivers are seamlessly integrated into the city's high-tech design.”
Adobe Firefly completely missed the mark on this one. For a start it isn't at night, the skyscrapers look like Playmobile toys and it took the parks a bit too literally. Meanwhile Midjourney captured the concept of a future city's upward sprawl perfectly.
Which model won?
Both did a good job in their own way, both have impressive fine-tuning and customization options — Adobe's are much easier to use than Midjourney — and both can create photorealistic images.
If your goal is to sell the image or use it in a commercial project where rights are a particularly sticky topic then Adobe Firefly is your only choice of the two, but for me the Midjourney images were all significantly better.
Both followed the prompt perfectly but Midjourney created more aesthetically exciting pictures based on the set criteria. There was more detail, depth and realism in every instance.
A key example of this is the future city where Midjourney generated something out of Sim City 2000 compared to Midjourney's year 3000 visual. That said, Adobe Firefly feels much more positive in its approach to image generation.
Midjourney wins.
