Start creating free

The "Emoji Test" on Popular AI Models

@user_google_961350@user_google_961350Β·Β·4 min readΒ·22 comments

@CaylaCatz recently put out a blog about using nothing but emojis for the prompt (https://budgetpixel.com/blog/emoji-mini-dictionary-part-1-people-activities); it looked like so much fun that I just HAD TO give it a try. Of course, being the difficult client I am, there is no way I would make things easy for the AI bots. So I gave them a string of emojis that do not seem to make sense together: πŸ‘΄πŸ½πŸͺ–πŸ‘˜πŸŽ€πŸ›ΌπŸ₯‚πŸ€.

First up was good old Chat GPT. To its credit, I have to say it did pretty darn well with these ransom objects--all seven are included in one neat package.

Illustration: The "Emoji Test" on Popular AI Models

Z Image Turbo, on the other hand, is clearly less inclined to indulge my chaotic mind. So out of "πŸ‘΄πŸ½πŸͺ–πŸ‘˜πŸŽ€πŸ›ΌπŸ₯‚πŸ€", I got this picture:

Illustration: The "Emoji Test" on Popular AI Models

A perfectly normal scene--and that's exactly the abnormal part of it. But my main question is: Where in the world did the kid come from?

Similarly, P-Image completely went off track--all I got from it is πŸ‘΄πŸ½ and none of the others parts of the prompt.

Illustration: The "Emoji Test" on Popular AI Models

Incidentally, I had never used P-Image prior to this, and I decided to give it a test run only because it's so cheap on BP. The introduction says that it is based on Alibaba's Qwen, so it really shouldn't be this bad.

Qwen itself, though, gave me an entirely different interpretation of the prompt. It obviously does not like the idea of a cross-dressing old man, since everything here has been rendered "gender-appropriate." As to the peaked cap in lieu of the helmet...well, it's peace time, after all.

Illustration: The "Emoji Test" on Popular AI Models

Here's Flux's interpretation of "πŸ‘΄πŸ½πŸͺ–πŸ‘˜πŸŽ€πŸ›ΌπŸ₯‚πŸ€." It substituted the kimono with a bath robe, but the whole thing is put together pretty well. Somehow, the pic kind of reminds me of the French foreign legion on parade. πŸ˜…

Illustration: The "Emoji Test" on Popular AI Models

Next, we have Minimax. Again everything is sort of there if we accept the bathrobe and the non-regulation basketball, but for some reason Minimax decided to make our fun-loving old man look rather sinister.

Illustration: The "Emoji Test" on Popular AI Models

Then it's Elon Musk's turn with his Grok. Clearly Grok tried to make sense out of these emojis that have little connection to one another and create a measure of normalcy out of the chaos, so it resorted to "divide and conquer." By distributing the objects among different characters, Grok managed to include all seven items listed in the prompt (if, again, we accept the peaked cap for the helmet) and come up with a scene that is actually quite "normal." A feather for Elon's hat for this demonstration of smarts!

Illustration: The "Emoji Test" on Popular AI Models

But of course, there is no way I'd let Grok off the hook that easily. To its credit, when I tell it to combine all seven emojis on the same character, Grok soon got the idea and asked me whether I wanted the scene to be more chaotic. Well, heck yeah! So, here's what we ended up with from Grok.

Illustration: The "Emoji Test" on Popular AI Models

Next, here's an interesting one by Midjourney--everything is sort of there, but not quite: old fella--check; roller skates--check; helmet--we got a military beret instead; kimono--a traditional Chinese outfit is close enough (geographically speaking); bow--a silk ribbon will do; champagne glass--a wine glass with whisky in it isn't too bad; basketball--I guess the stop watch is from the same department.

Illustration: The "Emoji Test" on Popular AI Models

The next pic is what Nano Banana came up with in response to the prompt "πŸ‘΄πŸ½πŸͺ–πŸ‘˜πŸŽ€πŸ›ΌπŸ₯‚πŸ€," and once again it is one of my favorite. In addition to putting all seven elements into the one coherent package, it actually put the sports equipment into action, unlike the other models, and made the whole thing work in a rather convincing scene. (BTW, did you know that roller skate basketball is an actual sport, and it even has a professional league with ten teams playing on national TV in India? I just found that out because of this...you learn something new every day!)

Illustration: The "Emoji Test" on Popular AI Models

Last but not least, the rendition by another of my favorite models, SeeDream. This was from SeeDream 5.0, and typically I like SeeDream 4.5, 4.0, or even 3.0 better than 5.0 because the pics made with 5.0 tend to be too stylized to look "real," but in this particular case I'd say the style it picked worked perfectly with the subject.

Illustration: The "Emoji Test" on Popular AI Models

Some of you may want to defend the AI models that did not do so well in this test by saying that those seven emojis are never meant to belong together. Oh yeah? Try this:

β€œCoach Ogawa promised the girls on the basketball team that, if they won the regional championship, he would come to the celebration on roller skates, wearing his wife’s kimono, and let them break a champagne bottle on his head.” 😊

Comments (22)

  • @CaylaCatzΒ·

    Videos do eat up credits. I'm about ready to try a few myself now that I've saved enough. I might be wrong on the credits but I don't think I am. I never hit the time thing myself because I use them up well before the date. I was told by someone who did hit a deadline that you get notified in your notifications when you are close to the expiration date. Since you're going to make videos, I doubt you'll hit an expiration date because it gets used up on a first come first served basis.

  • @user_google_961350Β·

    Thanks! Nano Banana has always been my favorite model among them all, and it did not disappoint this time either. My other favorite would be SeeDream, especially Version 4.5. It doesn't quite follow the prompts as well as Nano Banana, and it doesn't always give a "real-looking" portrayal, but it's a lot less restrictive with the subject matters. Lately, though, I've been really pleased with Qwen 2.0 Pro--for the same reason as with SeeDream.

  • @user_google_961350Β·

    My pleasure--obviously. 😁 My favorites have long been Nano Banana and SeeDream, so I'm glad they both did all right on the emoji test. GPT used to have a tendency to make all scenes somewhat dark, gloomy and brownish in color, but 2.0 seems to be different--although I haven't really played with it enough to get a good handle on this.

  • @SealineΒ·

    Yes! 1.0 had that distinctive look you described. GPT 1.5 fixed that problem, and I used for a long time. Now I find I mostly have shifted to 2.0 but 1.5 still does a phenomanal job and sometimes better. I have a strong dislike for the seedream girl look with 4.5 so mostly avoid unless editing an image to NSFW in i2i.

  • @user_google_961350Β·

    I don't think these credits expire (they land in the "extra" category on my account). I've had a few thousand of them sitting there for a while, and with the latest reinforcement I'll be able to play with SeeDance or the other video generators--they eat up credits like corn flakes.

Sign in to read all 22 comments and add yours.

"Guess the word" game

"Guess the word" game

I was writing a comment and remembered this puzzle game. The object is to reflect a word in an image, no letters involved. Here is my first puzzle to get the ball…

@DavidP@DavidPΒ·Β·2 min read
Design Challenge and Beyond - Part 2

Design Challenge and Beyond - Part 2

First, I appreciate all of you who took the time click this. I don't promote my blogs or tag a bunch of people so I would really appreciate it if you take the time to…

@Sealine@SealineΒ·Β·4 min read
THE RETURN OF THE PRIMORDIAL ONE

THE RETURN OF THE PRIMORDIAL ONE

Before the first star illuminated the darkness, before time acquired meaning, and before the universe gave birth to life, there was something that existed beyond the…

@archminalabib@archminalabibΒ·Β·7 min read
Design Challenge and Beyond

Design Challenge and Beyond

Inspired by @germancowboy blog where he discussed images extending beyond the Design Challenges at BudgetPixel. First, I appreciate all of you who took the time click…

@Sealine@SealineΒ·Β·4 min read

← All posts on the BudgetPixel blog