I all know that AI volumes, large language models, multimodal volumes, video volumes, and everything is rolled anyway.
AI graphics is one of the most mature modalities alongside large language models.
That's even more rolled up to fly.
A few days ago, Meta, a-stirring stick, officially launched their AI drawing model, called Meta Imagine, which is this thing.
Website: https://imagine.meta.com/ (The requirements for magic are relatively high, find a clean node in the United States)
The most important thing is that he is free.
Do you think he's a-stirring stick...
But Meta does have this confidence.,After all, you can't stand him Cardo.。。
A person who doesn't do cloud services, makes so many cards... Even the hoarding of H100 surpassed Microsoft.
What do you say he wants to do...
Of course, I'm here to be a-stirring stick (laughs.)
Of course,Actually, two days ago, I wanted to evaluate Meta's free AI drawing according to my system.,But because of PIKA1.0,There's really no time.,This Saturday afternoon I freed up my hands to have a good time.。
I also want to take this opportunity to compare and evaluate the four relatively large AI drawing models in my mind:
Meta Imagine,Midjourney,Adobe Firefly,Dalle。
The reason why SDXL is not put in it is because it is an open source model after all, and it is played by the fine-tuning and ecology of the follow-up gods, and the quality of the native is indeed a little bit worse...
So let's mainly compare these four large models.
I will evaluate from the four dimensions of detail quality, aesthetics (composition color, etc.), style diversity, and semantic understanding, and each dimension has 3 prompts, and at the same time, I will roll 3 times in the AI drawing model for each prompt, and take the most representative image to minimize bias.
At the same time, in order to have a final overall visual score for everyone to look more intuitive, I will score it. In each case, the first place is worth 4 points, the second is worth 3 points, the third is worth 2 points, and the last place is 1 point, and the sum is calculated last.
OK, let's get started.
1. Quality of details
It mainly tests the ability of AI drawing to express details, such as the texture of the character's facial skin, the details of the fabric texture, the details of the subtle elements of the scene, etc., which is a very important consideration for the accuracy of the model and the quality of the output.
Prompt1：Portrait of a 2000s blonde woman posing on a sports car, white wired headphones, expressionless, 2000s hairstyle, 2000s fashion, sun rays, light teal and amber,Cinestill 50D
Portrait of a 2000s blonde posing in a sports car, white wired headphones, expressionless, 2000s hairstyle, 2000s fashion, sun rays, light cyan and amber, Cinestill 50D
Obviously, it can be seen that Adobe has the best skin texture and clothing texture of the characters, followed by Meta and MJ, Dalle3 is the worst; There are problems with the details of the headphones, Dalle3 takes advantage of all the bugs, Meta doesn't draw it for you directly, and the details of the background are almost the same.
Adobe:4,MJ:3,Meta:2,Dalle:1。
-
Prompt2：Amazing photo of golden retriever chasing tennis ball underwater, close-up portrait
Amazing photo of a golden retriever chasing a tennis ball underwater, close-up portrait
Meta is the best overall, MJ secondly, the details that were wet with water are drawn, the details on Adobe Dog are a little less, Dalle3 is still pulled, and the details of the water bubbles collapse.
Meta:4,MJ:3,Adobe:2,Dalle:1。
-
Prompt3：A girl with a bunny sitting and smiling in 1970s fashion in a field of flowers
A girl with a bunny, dressed in 1970s fashion, sits among flowers and smiles
MJ wins, flowers and rabbits, hair details are basically nothing to pick, Adobe has enough details but pants are broken, Meta's facial skin texture is very uncomfortable, Dalle basically has nothing, a sense of oil painting.
MJ:4,Adobe:3,Meta:2,Dalle:1。
In terms of quality of detail, the overall score is as follows:
2. Aesthetics
It mainly tests the aesthetic ability of AI drawing, whether a picture is good or not, whether it is beautiful or ugly, in addition to details, more needs to look at the aesthetic ability of the model, such as composition, color, light and shadow, etc., the aesthetic is strong, and the picture is good-looking.
Prompt1：Product shot of juicy burger, artisan, rustic, food photography, delicious, close-up
Product shot of juicy burger, artisan, rustic, food photography, delicious, close-up
A picture that emphasizes aesthetics very much,Meta's color is almost impossible to see,Make people have no appetite,Dalle's composition is a big problem.,The background is too messy.,The two bottles are the same as the door god.,MJ doesn't have a composition, just a big subject.,Adobe wins.。
Adobe:4,MJ:3,Dalle:2,Meta:1。
-
Prompt2：Dungeons and Dragons, Close up of a fire breathing flying dragon, cinematic shot
Dungeons & Dragons, close-ups of fire-breathing dragons, movie shots
Close-up,Very emphasis on composition,There is also the contrast of light and shadow between fire and dragon body,Meta's dragon is extremely stupid,All the other dragon eyes will also glow to emphasize,It's really an eye.,Color and composition are not very good.,The overall best is Adobe.,Color and composition are great.,Next is MJ.,Dalle again.,The composition is almost meaningful.,It's too left.,The second is Meta.。
Adobe:4,MJ:3,Dalle:2,Meta:1。
-
Prompt：Diagonal Shot. Constantinople, 1453, masked sorceress, in the style of biblical drama, movie scene, low saturation, muted colors, extreme detail, 8K
Diagonal shooting. Constantinople, 1453, The Masked Witch, Biblical Drama Style, Movie Scene, Low Saturation, Pastel Colors, Extreme Detailing, 8K
MJ's composition and color texture basically exploded, Adobe didn't understand my final whitening aesthetic at all, and Dalle's composition was also strange.
MJ:4,Meta:3,Adobe:2,Dalle:1。
Aesthetically, the total score is as follows:
3. Diversity of styles
Mainly test the tolerance of AI drawing for style, such as Pixar style, Ghibli style, origami art, etc., theoretically it is necessary to use hundreds of art styles on a large scale to test the success rate, but I personally have limited energy, so simply test 3 slightly more common but not so bad street art styles.
Prompt1：an anime illustration of a samurai girl carrying a ninja sword, in the style of ethereal brushstrokes, ink painting, dark white and dark gray, fluid formation
Animated illustration of a samurai girl holding a ninja sword, ethereal style, ink painting, dark white and dark gray, fluid formation
In the ink painting section, the charm is still MJ and dalle, Meta's brushstrokes are very weird, not coherent and intermittent at all, and Adobe paints it like a Japanese comic.
MJ:4,Dalle:3,Meta:2,Adobe:1。
-
Prompt2：small boy looking out of his bedroom window into a cyberpunk world, pixelated, 8 bit style
Little boy looking out of his bedroom window at the cyberpunk world, pixelated, 8-bit style
8bit pixel art + cyberpunk, Adobe and Dalle have drawn this style, Meta is a little worse, MJ is not drawn at all. When it comes to pixelation, Adobe is really the best.
Adobe:4,Dalle:3,Meta:2,MJ:1。
-
Prompt3：Colorful logo of a French restaurant called "Khazix" with a flying seagull
A colorful sign of a French restaurant called "Khazix" with a seagull in flight
In terms of making Logo,Dalle3The precise text is indeed unique at present,No one compares,Logo design,Dalle is the strongest,MJ is second,Adobe is ordinary,Meta's graphics and details are simply sparse。
Dalle:4,MJ:3,Adobe:2,Meta:1。
In terms of stylistic diversity, the total score is as follows:
4. Semantic understanding
It mainly tests the ability of AI graphics to understand complex semantics, whether the text content can be clearly expressed and the quality of the generated images can be guaranteed.
Prompt1：A cup of coffee sitting on a table in front of a window; outside the window is a futuristic city; a futuristic monorail can be seen close by; many lush plants around; shot from ground floor; clouds above
A cup of coffee on a table in front of the window, a futuristic city outside the window, a futuristic monorail nearby, many lush plants around, taken from the ground floor, with clouds above
MJ crashed, and was the only one who didn't draw the train, Adobe drew the train but the track had a bug, Meta drew it but it was messy, and Dalle was perfect.
Dalle:4,Meta:3,Adobe:2,MJ:1。
-
Editorial photography of astronaut cooking Christmas colorful chocolate honey cookies on spaceship, Christmas honey cookies floating around astronaut, no gravity, in spaceship, levitated
Editorial photo of astronauts cooking Christmas colorful chocolate honey cookies on a spaceship, Christmas honey cookies floating around astronauts, without gravity, in a spaceship, levitating
Dalle beat the audience.,The only one who understands Christmas、Color elements.,Adobe is making cookies but doesn't have these elements.,MJ looks good but it's about to blow itself up.,Cookies aren't being made.,Meta's cookies aren't floating.。。。
Dalle:4,Adobe:3,Meta:2,MJ:1。
-
Prompt3：Shot diagonally. Cinematic shot of several astronauts in the space station, surrounding a chromium metal water droplet suspended in the air, the surface of the water droplet can reflect everything like a mirror, indoor scene
Diagonal shooting. Cinematic footage of several astronauts in the space station, around a chrome water droplet suspended in the air, the surface of which reflects everything like a mirror, an indoor scene
When I did "The Three-Body Problem" before, there was a sinkhole lens, and the chrome metal droplets that the mirror can reflect everything are not understood by few AIs.,Dalle is not the king of semantics.,Adobe understands it as a drop of water dripping from the sky.,Meta and MJ don't know what they're playing.。。。
Dalle:4,Adobe:3,MJ:2,Meta:1。
In terms of semantic comprehension, the total score is as follows:
Write at the end
After the evaluation of the four dimensions, we should be able to have a general understanding of these large models.
But to make it more intuitive, let's make a radar chart.
细节质量方面，MJ > Adobe > Meta > Dalle。
审美方面,MJ = Adobe > Meta = From。
In terms of stylistic diversity, Dalle > Adobe = MJ > Meta.
语义理解方面，Dalle > Adobe > Meta > MJ。
On the whole, at present, Adobe is the most bucket, the others are MJ, and then Meta, and Dalle is too biased.
Although only 20 prompts were released, I ran in the back for nearly 14 hours and measured more than 300 pictures. I'm about to throw up.
I hope this evaluation can lead you to some understanding of AI drawing synthesis.