news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

let's all play this game.

after openai “strawberry”, there is another mysterious modelblueberrycome to "dominate the list".

in the wensheng diagram model rankings, two "unclaimed" modelsblueberry_0andblueberry_1it beats well-known models such as flux.1, ideogram v2, and midjourney v6.1, becoming the first and second largest model on the latest list.

some of the generated effects be like, really good.

but this time, it may not be openai that crosses the border to cause trouble. the company behind it is more likely to be the black forest laboratory that created flux.

why?

let’s take a look at this set of pk. the bearded man generated by blueberry_0 is almost identical to the one produced by flux pro.

some people speculate that it is a new version of flux?

we also found a similar situation during actual testing. under the same prompt word, blueberry_1 generated images that were very similar to flux.1[pro] from composition, light to details.

some netizens commented that this is really interesting.

the company puts a new model on the model competition platform before releasing it, and most of the time users can guess which company it comes from.

……

this is almost a form of free advertising. and it can attract target users.

understand prompt words more accurately

you can now "encounter" two blueberry models by randomly matching them in the arena.

some characteristics of the blueberry model can be found from the actual generation effect.

for example, compared with midjourney v6.1, blueberry_1 is better in details and authenticity.

the characters generated by midjourney have unnatural hands, their eyes are looking at the table instead of the cash register, and the overalls have strange buttons.

relatively speaking, blueberry_1 is more coordinated overall and has more natural light.

compared with stable diffusion xl 1.0, the realism is better.

compared with flux.1[dev], blueberry_0 has a more accurate understanding of prompt words and a stronger ability to handle details.

the valley mentioned in the prompt is full of mysterious species, but flux.1[dev] only drew one.

but there are also some cases where blueberry_0's performance is not outstanding. who you think is better may be a matter of opinion.

in addition, some netizens found that the blueberry model performed poorly in generating text in images.

from a data point of view, the gap between the two blueberry models is not large. judging from the two dimensions of production quality and speed, they are also in the first echelon.

relatively speaking, the generation quality of blueberry_0 is a little higher. the following results are based on the evaluation of the model after pking more than 10,000 times in the arena.

the generation speed is also faster than blueberry_0.

in terms of arena winning rate, blueberry_0 ranks first, flux.1[pro] is second, and blueberry_1 is third.

netizen: is it openai or black forest?

no one has claimed the "blueberry" model yet, and speculation among netizens continues.

some people feel that it still has many problems with details and is unlikely to be an openai model.

for example, two door handles appeared in the cabin.

there is a problem with the blocked arms and hands of the character in the picture below.

but some people say that openai employees recently revealed that there will be new models, and it may be released quietly by openai.

of course, some people question the "blueberry" model's artificially high scores, and in some cases its performance is not as good as the dall·e 3.

i feel like the results are rigged so that they can show these results to investors. "blueberry" is obviously based on "strawberry", and the "strawberry" model is very suspicious.

what do you think?

now you can also go to the arena to test the effect of the "blueberry" model.