How good is Midjourney at different languages?
I took three different prompts and used Google Translate to translate them into several different languages before feeding them back to Midjourney (v6, same seed for all).
Let's see how Midjourney fared with this challenge!
Experiment 1: Samurai & Fox
The first prompt, in English is:
"samurai, fox, inkwash, splatter, color on black background"
I graded each language based on how well its images fulfilled each aspect of the prompt. If all four images had the aspect, it got a 1. If only some images had the aspect, it received the appropriate decimal.
Samurai | Fox | Inkwash | Splatter | Color/BG | Total | |
French | 1 | 1 | 1 | .75 | 1 | 4.75 |
German | 1 | 0 | .75 | .5 | 1 | 3.25 |
Finnish | 1 | 0 | 1 | .5 | 0 | 2.5 |
Spanish | 1 | 0 | .5 | .75 | .5 | 2.75 |
Italian | 1 | 1* | 0 | 0 | 1 | 3 |
Norwegian | 1 | 0 | 1 | 1 | 0 | 3 |
Irish | 1 | 0 | 1 | 1 | 0 | 3 |
Czech | 1 | 0 | 1 | .75 | 0 | 2.75 |
* It looks more like wolves to me, but I'm going to give it the benefit of the doubt that it's actually foxes.
1. French
2. German
3. Italian, Norwegian, Irish
4. Spanish, Czech
5. Finnish
As you can see above, "samurai" came through nearly universally thanks to it being a borrowed term and therefore similar in most languages.
"Fox" was a much harder token. Only French and (possibly) Italian had a true fox, but amusingly in Spanish, we can see the character "Zorro" filter through.
Most languages got the inkwash/splatter style right, but I can't help but if it's due to an association with "samurai" rather than any true understanding.
Non-Roman Alphabet Languages
You might have noticed that all the languages above use the Roman Alphabet. I did attempt to use languages that didn't. Let's check out those results:
As you can see, there are two versions of each of these language images: one in their own script and one that is in the Roman alphabet.
When I initially began generating in the non-Roman alphabets, I noticed that the results had nothing to do with the prompt and everything to do with the art style associated with that language.
Did the Romanized versions do any better though? I would argue no.
In the languages where the word "samurai" was close to the English spelling and had no accent marks (Hindi and Georgian), it came through.
You'll also notice that the foxes did come through for the Romanized Japanese prompt. I believe this is because "kitsune" is a popular enough word outside of Japan that Midjourney recognized it.
These were the only places where elements of the prompt worked.
Because these language simply revert to the language's style, I decided not to test them with the other prompts; however, if you are looking to add one of these styles to an image, using words from these languages may be an option.
Experiment 2: Woman & Raven
Our next prompt, in English, is:
"woman, raven, glowing magic, high fantasy epic, detailed comic book art"
For the rating, I've chosen to skip "high fantasy epic" because all of them successfully achieved that feel to me. Let's look at the other aspects.
Woman | Raven | Glowing Magic | Comic Art | Total | |
French | 1 | 1 | .25 | 1 | 3.75 |
German | 1 | 0 | 1 | .5 | 2.5 |
Finnish | 1 | 0 | .5 | 0 | 1.5 |
Spanish | 1 | 0 | .75 | 1 | 2.75 |
Italian | 1 | .5 | 0 | .75 | 2.25 |
Norwegian | 1 | .25 | 0 | 0 | 1.25 |
Irish | .25 | 0 | 1 | 0 | 1.25 |
Czech | 1 | 0 | .25 | 1 | 2.25 |
1. French
2. Spanish
3. German
4. Italian, Czech
5. Finnish
6. Norwegian, Irish
The raven was missed in most instances, but woman was nearly universal -- except in Irish where Midjourney was apparently confused by the translation "bean." It's possible Midjourney's strong preference for women was a factor here.
"Glowing magic" was well re-enforced both within the token and the token "high fantasy epic"; however, I was surprised how many images were in a comic-like style.
Experiment 3: Clock & Flower
The last prompt, in English, is:
"artistic double exposure photograph, clock, colorful flowers"
For this one we're dividing the first part into two: double exposure and photograph.
Double Ex. | Photograph | Clock | Flowers | Total | |
French | .5 | 1 | 1 | 1 | 3.5 |
German | 0 | 1 | 1 | 1 | 3 |
Finnish | 0 | 0 | 0 | .25 | .25 |
Spanish | 0 | 1 | 1 | 1 | 3 |
Italian | 0 | 1 | 1 | 1 | 3 |
Norwegian | 0 | 1 | 0 | 1 | 2 |
Irish | 0 | 0 | 0 | 0 | 0 |
Czech | 0 | 1 | .5 | 1 | 2.5 |
1. French
2. German, Spanish, Italian
3. Czech
4. Norwegian
5. Finnish
6. Irish
"Double exposure" was a difficult term, but somehow it shown through in half of the French images.
Languages with a close spelling to "photograph" ("photographie," "fotografia," "fotografie") came out as photographs. German did not have a close word, but also came out as photographs.
Unfortunately Finnish and Irish seem to have reverted to something like their traditional style. But Norwegian and Czech at least got the flowers (and even a clock in one case for Czech!)
Conclusion
Samurai & Fox | Woman & Raven | Clock & Flowers | Total | |
French | 4.75 | 3.75 | 3.5 | 12 |
German | 3.25 | 2.5 | 3 | 8.75 |
Finnish | 2.5 | 1.5 | .25 | 4.25 |
Spanish | 2.75 | 2.75 | 3 | 8.5 |
Italian | 3 | 2.25 | 3 | 8.25 |
Norwegian | 3 | 1.25 | 2 | 6.25 |
Irish | 3 | 1.25 | 0 | 4.25 |
Czech | 2.75 | 2.25 | 2.5 | 7.5 |
1. French
2. German
3. Spanish
4. Italian
5. Czech
6. Norwegian
7. Finnish, Irish
If you wish to use a different language with Midjourney, the safest out of those tested is French, followed by German, then by other romance languages.
That said, from what I understand, Midjourney has improved the language understanding through its versions and it is sure to continue to improve. It makes sense that as an American company, most of its data is in English, and some languages simply will not have as many captioned images available.
Until the language comprehension improves, using non-English languages may be a way to introduce either randomness or traditional styles into an image.
Do you speak any of the languages tested here? Were the prompts good translations?
What have your experiences with prompting in languages besides English been like?
Let us know in the comments below or in this thread on X!
Comments