Can AI Eat Ramen?

2024-11-212025-2-16

Anime illustration of a boy eating ramen with his hands.png (1024×1024)

Hot! Watch out for burns.

Introduction

Hello, this is Easygoing.

Today, I’m introducing the hotly discussed AI Ramen Problem in the AI illustration world.

What is the AI Ramen Problem?

In the early days of AI image generation (around October 2022), it became a topic of discussion that "AI struggles to depict people eating ramen properly!"

#AI Ramen Challenge Drawings, Best Fan Art on pixiv, Japan

This issue was brought up about two years ago, but how does it stand today?

Stable Diffusion XL

First, let’s try with Stable Diffusion XL, the tool I primarily use (from July 2023 onwards).

A boy struggling to eat ramen properly.png (1024×1024)

A boy eating cup ramen in his apartment at night.png (1024×1024)

A boy determined to eat ramen correctly despite his struggles.png (1024×1024) — "Don't look at me!"

Stable Diffusion XL struggles to depict someone eating ramen properly.

Udon and Soba Face Similar Issues

It’s not just ramen. The same problem occurs with udon and soba noodles.

A woman awkwardly holding four chopsticks.png (1024×1024)

Noodles flying in odd directions.png (1024×1024)

The Core Issue: Bias in Training Data

Image generation AI learns from a massive dataset of images. One of the most commonly used datasets, such as LAION-5B, includes billions of images sourced from Pinterest, WordPress, Blogger, Flickr (Yahoo-related), and more.

AI learns to associate weights between images and related text, generating outputs that align closely with the input text.

Scarcity of Specific Images

Even with a vast dataset, narrowing conditions reduces the number of relevant images.

All images ⊇ Food-related images ⊇ Images of chopstick use ⊇ Images involving noodles

Images of people eating ramen with chopsticks become scarce, and capturing the precise moment of slurping ramen is almost non-existent due to cultural etiquette.

When AI cannot learn from accurate examples, it’s more likely to produce strange outputs.

How Do Other AIs Perform?

Let’s test other AIs, like Adobe Firefly Image3 (from April 2024).

A muscular man eating ramen with chopsticks in both hands.png (2048×2048) — A smiling foreign bodybuilder eating ramen with chopsticks at a Japanese night stall. Food meets culture.

Another muscular man eating ramen with chopsticks—on closer inspection, the grip seems odd.png (2048×2048)

Firefly performs much better, though the chopstick handling is still slightly off. Unlike LAION-5B, Firefly uses datasets from copyright-free images and Adobe Stock, providing high-quality training data.

For instance, if a user uploaded a ramen-eating photo series to Adobe Stock, AI could learn effectively from such quality inputs.

Flux.1’s Performance

How about testing Flux.1, the latest AI released in August 2024?

Anime illustration of a girl with orange hair eating ramen noodles skillfully with chopsticks.png (1440×1440)

Anime illustration of a boy with black hair eating ramen noodles skillfully with chopsticks.png (1440×1440)

Flux.1 handled the task impressively! While its training data isn’t disclosed, it might include proprietary datasets supplementing LAION-5B.

Imagen 3 by Google

Google Imagen, trained using Google Photos, boasts immense data volume.

Google Imagen's anime illustration of a brown-haired boy skillfully eating ramen noodles with chopsticks.png (1024×1024)

Google Imagen's anime illustration of a dark-haired boy eating ramen noodles skillfully with chopsticks.png (1024×1024)

The result? Absolutely perfect.

AI Image Generators Continue to Evolve

AI image generators are constantly improving over time.

Below is a timeline of major advancements in image-generative AI:

gantt
title History of Image Generative AIs
dateFormat YYYY-MM-DD
tickInterval 12month
section Open AI
Dall-E :aa1, 2021-01-05, 2022-09-28
Dall-E2 :aa2, after aa1, 2023-08-20
Dall-E3 :aa3, after aa2, 2024-10-31
section Midjourney Inc.
Midjourney V1 :done, ab1, 2022-02-01, 2022-04-12
Midjourney V2 :done, ab2, after ab1, 2022-07-25
Midjourney V3 :done, ab3, after ab2, 2022-11-05
Midjourney V4 :done, ab4, after ab3, 2023-03-15
Midjourney V5 :done, ab5, after ab4, 2023-05-03
Midjourney V5.1 :done, ab6, after ab5, 2023-06-23
Midjourney V5.2 :done, ab7, after ab6, 2023-12-20
Midjourney V6 :done, ab8, after ab7, 2024-07-30
Midjourney V6.1 :done, ab9, after ab8, 2024-10-31
Nijijourney :done, n1, 2022-11-07, 2024-10-31
section Stability AI<br>Stable Diffusion 1
Stable Diffusion 1.1 :a1, 2022-08-22, 2022-08-28
Stable Diffusion 1.2 :a2, after a1, 2022-09-03
Stable Diffusion 1.3 :a3, after a2, 2022-09-13
Stable Diffusion 1.4 :a4, after a3, 2022-10-11
Stable Diffusion 1.5 :a5, after a4, 2024-10-31
section Stable Diffusion 2
Stable Diffusion 2.0 :b1, 2022-11-24, 2022-12-07
Stable Diffusion 2.1 :b2, after b1, 2024-10-31
section Stable Diffusion XL        
Stable Diffusion XL 0.9 :c1, 2023-04-12, 2023-07-26 
Stable Diffusion XL 1.0 :c2, after c1, 2024-10-31
section Stable Diffusion 3                
Stable Diffusion 3 :d1, 2024-06-12, 2024-10-22
Stable Diffusion 3.5 :d2, 2024-10-22, 2024-10-31
section Anlantan LLC
Novel AI Diffusion :crit, e1, 2022-10-03, 2023-04-12
Novel AI Diffusion Anime V2 :crit, e2, after e1, 2023-11-06
Novel AI Diffusion Anime V3 :crit, e3, after e2, 2024-10-31
section Leonardo AI
Leonardo AI :done, f1, 2023-09-20, 2024-10-31
section Kuaishou Technology
Kolors :g1, 2024-07-06, 2024-10-31
section Black Forest Lab
Flux.1 :h1, 2024-08-01, 2024-10-31
Flux.1.1 :h2, 2024-10-03, 2024-10-31
section Adobe
Firefly :crit, ac1, 2023-03-14, 2023-06-05
Firefly 2 :crit, ac2, after ac1, 2024-04-23
Firefly 3 :crit, ac3, after ac2, 2024-10-31
section Google
Imagen :i1, 2022-05-23, 2023-05-09
Imagen 2 :i2, after i1, 2024-08-15
Imagen 3 :i3, after i2, 2024-10-31
section Ideogram
Ideogram :done, j1, 2023-08-29, 2024-08-20
Ideogram 2 :done, j2, after j1, 2024-10-31
section Meta
Imagine :k1, 2023-12-06, 2024-10-31 
section Fal.ai           
AuraFlow 0.1 :crit, l1, 2024-07-12, 2024-07-26
AuraFlow 0.2 :crit, l2, after l1, 2024-08-14
AuraFlow 0.3 :crit, l3, after l2, 2024-10-31

[AI Image Generation 3] Still More AI Image Generators! Are They Available for Commercial Use? | AI image journey

Explore More AI Generators and Commercial Usability

Today, we tackled the AI Ramen Problem. While AI still struggles with certain scenes, these issues are bound to improve over time.

Thank you for reading until the end!