Nvidia AI experts measured: Currently MidJourney beats Adobe Firefly
In the past two days, many people have begun to get Adobe Firefly test qualifications.
Although it has just entered in late March, as a new AI tool from an authoritative company in the design industry , its generation effect is naturally expected to be high .
No, Nvidia AI research scientist Jim Fan took the lead in giving the actual measurement:
Can Adobe compete with self-funded 11-person team MidJourney?
Well, Adobe Firefly got the hang of it ...
Do not believe you see:
Adobe Firefly vs. MidJourney
Jim Fan tested a total of 8 rounds.
Deadpool strikes a wide-angle pose on top of a car outside an apartment complex in the UK.
Look at the generated results of the two, Firefly on the left and MidJourney on the right (the same below):
Ah, Firefly, are you sure the one on the left is Spiderman, not the real Pigman?
A photo captures Super Mario posing on a dark street with a big reflection in a puddle on the ground, a slight mist in the air, and you know something is about to happen and people are looking forward to it. It's like a scene from a Hollywood movie, high production value, anthropomorphic shots.
Not to mention the image of Super Mario, I feel that Firefly is also inferior in the depiction of the background and atmosphere.
Exactly the same prompt as the second round, but replace Super Mario with Pikachu.
Apart from the image of Pikachu, MidJourney is better able to capture the atmosphere, and Firefly can't even perform well in dark streets this time.
Let's not adjust the IP image, let's see who is better at character photography in this round.
This modern street style photo from Fujifilm captures a morose morning in Tokyo with a large group of commuters walking the streets. The attention-grabbing faces in the crowd wear a mix of Japanese streetwear and business attire, but it's the woman in red that grabs your attention. The warm glow of the sunrise blooms through the mist, while the sharp sunlight and careful composition emphasize the mood of the crowd and the elegance of the woman in red.
There are two completely different presentation methods, but Firefly is really a bit distorted.
This round is a little more sci-fi.
Abstract fractal artwork, circular mosaic urban architecture made of millions of 3D alien shaped blocks, the city is located in space, Jupiter can be seen in the background, as seen from a spaceship.
I feel that Firefly's understanding of "circle" is too superficial, and obviously he doesn't understand what "as seen from a spaceship" is.
Now to generate some icon designs for mobile apps, does Firefly know what this is?
App icon design: iOS, sci-fi planetary landscape, skeomorphic style; visual technique: use shallow depth of field, high exposure, soft focus to emphasize beauty; equipment: shot with high-resolution camera; editing: enhanced with color correction and sharpening; Platforms: Dribble, Behance
Apparently it doesn't know yet.
Seventh round :
Newspaper/TV/Journalism style photo, medium panorama, afga vista still, of a young Jamaican woman with white braids standing in the desert, wearing a vintage sequined Gucci white dress.
This round is mainly to look at hand generation, which AI is least good at.
At first glance, both are okay, but if you take a closer look, you will find that Firefly has given an image of a man in disguise, with disabled hands, and the background still loses a bit compared to MidJourney. However, the appearance of the "white braid" of the two is the eldest brother, not the second brother.
Red Ferrari F40 among dandelions by Lake Seealpsee, shot with Fujifilm Velvia 50, high detail, high contrast 4k
At first glance, it’s okay, but the background of Firefly is still fake, but Firefly said that it only draws dandelions when it draws dandelions, and MidJourney adds some other flowers to it, which is slightly off topic.
That's all for the test. Overall, Firefly is really a complete failure. There is a kind of rush for elementary school students VS college students. Some netizens think that the level of Firefly is also the V1 version of MidJourney.
However, as Jim Fan said, this situation can be forgiven, after all:
(1) These hints have been optimized a lot in MidJourney, so it's not fair to directly compare them with Firefly, which just "debuted".
However, Jim Fan also tried some "variants", but he still couldn't get better results. It may also be related to the fact that the prompt words he wrote were not very good. The prompt words of the above tests basically came from netizens.
(2) Most importantly, Firefly claims that it was trained only on Adobe Stock and fully licensed images.
This is a big disadvantage, but it is an important advantage for users who must ensure the copyright of the output image. Because these people are likely more willing to sacrifice quality for legitimacy.
In response to this topic, Jim Fan further expressed his views:
IMHO, neither MidJourney's strategy (little focus on artists and copyright) nor Adobe's strategy (weakening training to protect intellectual property) is suboptimal in the long run.
There are also some good works
In fact, if you search the Internet casually, you can still see some good works generated by Firefly without such comparison.
It can be regarded as salvaging some situations.
For example this:
Old Albanian woman walking on the cobblestone road.
And the animation effect added with it:
I can feel that Firefly is not good at realism, and this kind of cartoon/hand-painted effects are still possible.
Midjourney update: can generate text with pictures
The mighty Midjourney continues to iterate.
The official recently released a command: /describe , which can describe a picture with 4 text results.
Netizens said one after another: We can use the description it gives to study how to write prompt words better.
However, directly using the description given by it to regenerate the picture will be somewhat different from the original picture, such as the "subway old man looking at the phone" meme tried by the design aesthetics blogger @Simon_阿文:
But soon some netizens gave a solution:
In addition to manually modifying the prompt words, feeding the original image to Midjourney during the secondary generation will basically make little difference.