[AI Image Generation Tool] Who Will Win!? A Thorough Verification of 13 Models on Pollo AI
The AI platform introduced in the article above, "Pollo AI", is equipped with many image generation AIs, with a total of 13 types! (As of June 2025)
In this article, I will introduce and compare the image generation AIs available on Pollo AI.
The generation theme is "Mitarashi Dango" 🍡
The 3rd, 4th, and 5th of every month are Mitarashi Dango Day (established by Yamazaki Baking Co., Ltd.). The date is based on a pun for "mi" (3rd), "tara" "shi" (4th), and "dan" "go" (5th).
The prompt is just a rough "Mitarashi Dango". It's a minor subject that might depend on whether it's included in the data the AI has learned, but that's exactly why it might produce interesting results!
It does appear! (Spoiler alert)
Credits are points that can be used on Pollo AI. Please consider 1 credit to be approximately 5 yen.
① Flux Schnell (1 credit)
Flux is a relatively new image generation AI model developed by Black Forest Labs. It is said to aim for performance exceeding Midjourney, and is characterized by high-quality image generation and high-speed processing.
A shop that looks like it sells "Mitarashi Dango" appeared, but is this also what you would call a hallucination? This is a bad example of what happens if you don't specify the prompt in detail. The atmosphere of the picture is good, though. The potted plant in front of the shop seems to be in the way.
1 credit is a nice price to generate with. Other models cost at least 5 credits.
② Flux Kontext Pro (12 credits)
Flux probably doesn't have "Mitarashi Dango" in its training data. Even if performance improves, it cannot generate things it doesn't know (it just quietly generates something else). The quality of the generated image is photorealistic and beautiful.
3. Imagen 4 (8 credits)
Looking at the first and second images, I was worried about what I would do if it didn't properly produce Mitarashi Dango, but Imagen 4 generated some delicious-looking Mitarashi Dango!
Six dumplings on a skewer is quite bold. The slight char also looks fragrant and delicious.
It feels a bit strange that they are on a bamboo noodle mat. Do some shops serve them that way?
4. Imagen 3 (10 credits)
A scene of drizzling the sauce. They are not on skewers.
Compared to Imagen 4, this might be a bit lackluster?
By the way, if you try to generate “Mitarashi Dango” directly using ImageFX, which is based on Imagen 3, it was not possible. It seems there is no training data for “Mitarashi Dango” itself.
Of course, if you specify it in detail with a prompt, you can generate it.
Close-up of traditional Japanese Mitarashi Dango – skewered rice dumplings glazed with shiny sweet soy sauce, placed on a ceramic plate, soft natural lighting, detailed texture of the sauce, shallow depth of field, appetizing food photography, high-resolution
5. GPT 4o (5 credits)
Delicious-looking “Mitarashi Dango” generated with three dumplings. The glaze on the sauce is realistic. The plate is also an appropriate size. Maybe the amount of sauce is a bit much?
6. DALL•E 3 (12 credits)
DALL·E 3, developed by OpenAI, the company famous for ChatGPT. The "Mitarashi Dango" generated with GPT-4o looked delicious, but is this a bit lackluster...? The tip of the skewer sticking out is also a bit scary. I thought DALL·E 3 was used when generating images with ChatGPT 4o, but is it slightly different? I don't really understand the fine details of the specifications.
⑦ Recraft v3 (10 credits)
Recraft v3 is an AI image generation and design tool developed and provided by Recraft, a British design tool production company. The competition among AI tools is fierce, and there are so many tools that it's hard to keep track of them all... That is precisely why I think it is important to use various tools on a platform like Pollo AI (promotional PR).
Compared to the Mitarashi sauce on the plate, the amount of sauce on the dumplings seems a bit low. The texture of the dumplings is very realistic, showing the coarseness of the rice grains.
⑧ Ideogram v2 Turbo (12 credits)
A Japanese garden and koi. The bridge railing is cut off in the middle. No matter how you look at it, this is not "Mitarashi Dango".
Ideogram is a tool developed by Ideogram.
Ideogram is an AI startup founded by former Google Brain researchers. It is said to have been created by experienced AI experts who were involved in the development of Google's image generation AI, such as "Imagen" and "Imagen Video".
The biggest feature of Ideogram lies in its ability to generate accurate text within images.
It is not utilizing that feature at all. Let's try it.
A close-up of traditional Mitarashi Dango (Japanese rice dumplings on skewers), with the kanji characters "御手洗団子" written clearly across the dumplings, glossy sweet soy glaze, placed on a ceramic dish with a Japanese aesthetic background. Warm lighting, high detail, realistic food photography.
As expected, it seems it was difficult to draw Japanese food with kanji...
⑨ Stable Diffusion 3 (10 credits)
Even with Stable Diffusion, which prides itself on customizability, it doesn't know what it doesn't know. It's problematic when a woman appears just because it doesn't know. Is this what Stable Diffusion imagines "Mitarashi Dango" to be? I don't really understand the generation principle of hallucinations.
I will regenerate using the prompt I used earlier in ImageFX.
Close-up of traditional Japanese Mitarashi Dango – skewered rice dumplings glazed with shiny sweet soy sauce, placed on a ceramic plate, soft natural lighting, detailed texture of the sauce, shallow depth of field, appetizing food photography, high-resolution
⑩ Flux Dev (5 credits)
⑪ Flux Dev Lora (5 credits)
⑫ Flux 1.1 Pro (8 credits)
⑬ Flux 1.1 Pro Ultra (12 credits)
I don't really understand the hallucinations in Flux! Whether it was a character imagined from the word "Mitarashi Dango," storyboards appeared, chibi characters appeared, and in the end, food that looked slightly unappetizing appeared.
It is true that it's because I didn't set the prompt properly, so here is a "Mitarashi Dango" to redeem its honor (with detailed prompt specifications)
This kind of food seems like it could actually exist. Even if it's not the traditional Japanese "Mitarashi Dango." If it looks delicious, that's fine! (?)
Summary
ChatGPT 4o is the winner! It can generate the most delicious-looking "Mitarashi Dango" without even specifying a detailed prompt!
Imagen 4 is the runner-up! It generates delicious-looking "Mitarashi Dango" without needing a detailed prompt.
Imagen 3 (ImageFX) takes 3rd place! Even without specifying detailed prompts, it looks reasonably delicious, and if you specify it in detail, it can generate it properly.
Bonus: "Mitarashi Nyanko" (Mitarashi Dango Cat) 🐈
The prompt is just "Mitarashi Nyanko"
Winner! Too cute! It has a great expression.
This is also a winner! The two models that have already learned about Mitarashi Dango are strong.
It's like a cat hot pot. Looks like a cat pudding.
I thought it would be amazing if I could generate a cat that looks like "Mitarashi Dango" just with the word "Mitarashi Nyanko," and when I tried it, it really did generate it, so I thought it was amazing after all.
"Mitarashi Nyanko" generated with ImageFX
The prompt is specified in detail (it's a secret)
Use Pollo AI easily on your smartphone!
It is convenient to have the app installed for times when you want to do a little work during your spare time while on the move, or when you want to use it for fun with friends or a partner. Of course, you can link your account.
Click here for the iOS app↓
Click here for the Android app↓
[Includes free credits] Finally
Pollo AI account registration can be done from here! (You will receive 10 credits!)
Let's have fun creating cool videos together!
Pollo AI still has many features that I haven't been able to introduce yet, so I will find another opportunity to introduce them.