Three months after the announcement, the popular Vincentian model Midjourney has finally launched its latest version. On December 21, local time, Midjourney announced on Discord the release of the beta version of its latest version, V6, which is currently in the alpha testing stage. Judging from a large number of examples from users, V6 is very good at handling realistic pictures and abstract paintings, and its effect is as good as that of designers and photographers.
Midjourney positions V6 as a major innovation. Its CEO DaVidHolz said that V6 is actually the third set of models trained from scratch on an AI super cluster. The entire development cycle lasted nine months. "The images generated by this set of models are far more realistic than any version we have released before." According to the official introduction, the main changes of V6 are better picture quality, stronger semantic understanding, the ability to embed text, accommodate more prompt words, higher coherence, and richer model knowledge.
According to user testing, V6 now supports prompt words longer than 350 characters, and can even understand subtle differences in punctuation and grammar. Judging from the images currently displayed by netizens, V6 is indeed a step up from the previous generation in terms of understanding and details such as light and shadow, composition, material, and color.
Use the same prompt to test V6 and V5.2, the contrast is very obvious (the above picture is generated by V6; the lower picture is generated by V5.2):
Key word: 1980s suspense movie, shot from above, a French butler in a black suit holding a candle in the corridor of a Victorian mansion
Main prompt: 1960s street style photo of a young woman wearing a green silk dress and pearl necklace sitting on a sailboat
Main cue word: Female operator wearing a high-collared silver operating suit from a retro 1940s science fiction movie
Key cue word: The neon sign at the corner bar says "Open Until Late"
Main cue word: Sunset reflection in rain puddle
Main cue word: A pot of stew, served with a wooden spoon
In terms of text generation, V6 can embed text in images more clearly and even specify its style.
Note: Coca-Cola original text: CocaCola
Restore the texture of sweaters, animal hair, and raindrops on windows
Handling of long text is also better
Product logo
Comparison of product design drawings with text from different tools
This performance improvement is expected to bring greater gains to the design and marketing industries. It is understood that some cross-border e-commerce practitioners have long used Wenshengtu large models to create product introduction pages and model display pictures. Midjourney is the most commonly used tool.
In addition, V6 can "paint hands". Previously, AI paintings have been criticized for being unrealistic, especially the details of characters' hands, which often appear deformed. But with the launch of V5, it has perfectly solved this technical problem and can even display the fingerprints and skin texture of the hand, achieving a leapfrog breakthrough in AI painting. Below are some hand drawings:
Currently, the V6 is missing some features found in the V5.2 model, including left-right balancing and zooming out, but Holz said these features will be implemented in subsequent updates to the V6.
V6 will not be the end of Midjourney. The product has been in iteration. The first version was launched in March 2022, and then quickly evolved to the current sixth version, updated every three months on average. In Midjourney's announcement, they said: V6's speed, image quality, coherence, prompt following, and text accuracy should improve in the coming weeks. V6beta announced its first update half an hour after its release, increasing the generation speed by 2.7 times.
Previously, the company also stated that future technology update directions include generating 3D and video. Holz predicts that it will be possible to generate content in real time at a high resolution of 30 frames per second, and by 2030, entire video games may be generated.
It is worth mentioning that founder David Holz allegedly rejected the olive branch offered by venture capitalists many times. In the past year, the number of Midjourney users on the Discord platform increased from 2 million to 17.67 million, with more than 100,000 users every day. One million people are online (as of press time), and the product has already launched a paid model. Users can choose from different packages and charge US$10 to US$120 per month. With a team of 40 employees, Midjourney successfully achieved an annual profit of US$200 million in September.