The new text-to-image models by Black Forest Labs are making waves since their inception last week, and posts and images eulogizing popular Midjourney started popping out right away. Let's take a look at a few generations to see which model drives better results!
My KPIs:
Prompt adherence (how much the image followed the prompt)
Image Quality
Creativity
X factor (that makes an image beautiful
Cherry picking method: A batch of four images on Midjourney, 1-4 on Flux Dev.
wide shot, long shot, full body of A girl dancing, with hair made of colorful melting ice cream, featuring sprinkles and a cherry on top. She is wearing pink sports hoes and She has a playful expression, with ice cream dripping slightly down her face. Bright, a car garage on fire in background grainy, shot on film
Flux Dev
Midjourney
Both images are stunning and hit the spot on all KPIs
wideshot, longshot, fullbody, a lady wearing a wide pink hat, spilling orange juice on a surprised zombie creature, in panic, running away , dark alley
Flux Dev
Both models did not fully adhere to the (wierd) prompt, but at least on Midj showed the zombie I requested. Also in terms of cinematic and X-factor Midj scores higher.
a photo of 2 birds playing pool in a night club, one of the birds is smoking, the other holds a drink
Flux Dev
Midjourney
If we ignore the fact the Flux's bird is holding its drink with it's third leg, the visual is stunning. Both images do not follow the prompt one to one but deliver strong results nonetheless.
Wide shot, a lady wearing a wide pink hat holding an orange juice, riding a zebra in savanna Africa
Flux Dev
Midjourney
Flux Dev has the upper hand here albeit Midj is more stylstic. In both models I had problems having the images show the full landscape as a wide shot.
a bowl of yogurt with the text "Boker Tov Shir" written in honey, top down view, a breakfast table | a graffiti on the streets of Manchester "I'm so sorry Why do you come here When you know it makes things hard for me? When you know, oh, why do you come?"
Flux Dev
Midjourney
Text is one of Flux's strong suits, but as you can see, Midjourney handles the task pretty well and is slightly more cinematic.
Pixel art of two girls enjoying ice cream on a vibrant street in Tel Aviv. The girls, with bright outfits, savor colorful ice cream cones. The background features charming buildings, trees, and a clear blue sky
Flux Dev
Midjoruney
Both models fail to deliver a decent pixel art image right out of the box, but Midjourney certainly tried harder.
A young witch with bright purple hair and a mischievous grin, riding a broomstick over a bustling, whimsical town, Below, candy-colored houses and quirky shops line winding streets, while enchanted creatures and magical beings go about their day, The sky is filled with sparkling stars and swirling clouds, creating an enchanting twilight scene, vibrant and full of life
Flux Dev
Midjourney
Both models scored high on all KPIs, despite ignoring the "while enchanted creatures and magical beings go about their day" part. Albeit, Midjourney's outputs seem more alive and energetic, and less trivial, despite the strange broom which can be fixed easily.
Verdict
I could go on generating for hours but had to conclude this post at some point :)
Both models are amazing and score high on all KPIs. No more 6-fingered hands (most of the time) and cherry-picking made easy (I generated no more than 4 images per prompt). But which model is better?
?
Well, I wouldn't cancel my Midjourney subscription right away. Comparing these two models, I think that Midjourney is still the best image model around. It's creative, fast, generates artistic results in high quality, and comes with a set of tools such as style reference and character reference that take some load off the prompting part of the generation process.
However, Flux (schnell & dev) shows HUGE potential. The quality of the images is already amazing, and prompt adherence is top-notch. It could be more creative and versatile, and definitely has a mile to go on the illustration part. But this is only the beginning for Flux. The fact that the weights are distributed to the public is a huge plus. Soon, plugins and extensions such as Controlnet and IPadapter (style reference) will come out, which will make Flux more controllable and versatile. Soon, we will also figure out how to train new concepts on the infrastructure of Flux, and then, I may wear my "cancel Mindjourny Subscrpition" T-shirt.
Comments