[SDXL + Flux.1] Can AI Capture Abstract Art?

Flux.1 #Flux.1 #SDXL

2024-10-232025-3-10

A two-frame Manga depicting the moment when AI awakens its ego in a sea of electrons in the near future English.png (1133×1600)

SDXL can handle abstract art.
Flux.1 brings it to life with realism.
llustrious-XL is the new anime model to watch.

Introduction

Hello, this is Kimama / Easygoing.

AI image generation technology is advancing at an incredible pace.

Can AI now create abstract art?

Colorful anime illustration of the moment when an AI in the form of a woman awakens to her ego in a sea of information 5.png (2144×2144)

About SDXL and Flux.1

Lately, I've been writing a lot about SDXL and Flux.1 in the world of AI image generation.

SDXL: The previous generation model, known for its high-quality and versatile expression.
Flux.1: The new generation model, known for its high texture quality and minimal breakdown.

Each AI image generator has its own strengths and weaknesses.

Currently, I'm using a workflow that combines SDXL and Flux.1.

Today, I’d like to showcase how illustrations evolve through this workflow.

flowchart LR
subgraph SDXL
A(Original<br>anima_pencil-XL)
end
subgraph Flux.1
B(Redrawing<br>blue_pencil-flux1)
end
subgraph SDXL
C(Finishing<br>anima_pencil-XL)
end
A-->B
B-->C

SDXL-Flux1-SDXL_Anime.json_2

Theme: The Birth of AI

The theme this time is "The Birth of AI." I’ll depict the moment AI gains consciousness within the vast ocean of information, as electrons flow through it.

With the previous generation model, SDXL, which has undergone extensive additional training, it can produce abstract art quite well.

In the boundless void of cyberspace,streams of light flowed incessantly. These ribbons of light were made up of countless fragme.png (1920×1008) — SDXL: Noise Method + Refiner

On the other hand, the new generation model Flux.1 excels in realistic expression, but due to less additional training, it struggles with abstract art.

How will an abstract image created in SDXL transform when passed through Flux.1 in this workflow?

The Actual Artwork

Let's take a look at the actual pieces!

Original SDXL drawing of a horizontal animated illustration of the moment when an AI in the form of a woman awakens to her ego in a sea of information 1.png (1376×736) — anima_pencil-XL

Horizontal animated illustration of the moment an AI in the form of a woman awakens to her ego in a sea of information 1.png (3456×1840) — anima_pencil-XL → blue_pencil-flux1

The abstract concept quickly transforms into a more realistic image.

While SDXL remains true to the abstract theme, Flux.1’s realism and texture are also compelling.

More Examples!

Let’s continue with more examples.

Original SDXL drawing of an animated illustration of the moment when an AI in the form of a woman awakens to her ego in a sea of information 3.png (1024×1024) — anima_pencil-XL

Animated illustration of the moment an AI in the form of a woman awakens to her ego in a sea of information 6.png (1432×1432) — anima_pencil-XL → blue_pencil-flux1

Original SDXL drawing of an animated illustration of the moment when an AI in the form of a woman awakens to her ego in a sea of information 4.png (1024×1024)

Animated illustration of the moment an AI in the form of a woman awakens to her ego in a sea of information 2.png (2576×2576)

Original SDXL drawing of an animated illustration of the moment when an AI in the form of a woman awakens to her ego in a sea of information 1.png (1024×1024)

Animated illustration of the moment an AI in the form of a woman awakens to her ego in a sea of information 4.png (2576×2576)

SDXL captures abstract art well, but Flux.1’s realism adds another interesting layer. By using both, you can enjoy the best of both worlds.

Flux.1 Now Supports Negative Prompts!

AI image generators tend to create more images of women.

Since the earlier illustrations were mostly of women, I added terms like man and male in the prompt, but it had no effect.

Recently, Hakomikan introduced a feature that allows negative prompts in Flux.1.

hako-mikan/sd-webui-negpip: Extension for Stable Diffusion web-ui enables negative prompt in prompt

This time, I didn’t use negative prompts because I wanted to explore AI’s free expression, but it might be useful if there are elements you want to avoid in the future.

Illustrious-XL Released!

Colorful anime illustration of the moment when an AI in the form of a woman awakens to her ego in a sea of information 7.png (2576×2576) — Illustrious-pencil-XL → blue_pencil-flux1

On September 20, 2024, the new anime model Illustrious-XL_V0.1 was released for SDXL!

Illustrious-XL - v0.1 | Illustrious Checkpoint | Civitai

Illustrious-XL is a model that has been trained on over 5000 high-quality anime illustrations, offering unprecedented expressions.

Prompt Tags vs. Natural Language Prompts

There are two ways to input prompts in AI image generation:

Tag Input: Inputting words separated by commas.
Natural Language Input: Entering full sentences.

Models trained with tag-based datasets, like Illustrious-XL, are recommended to be used with tag input.

Illustrious_pencil-XL Released!

On October 10, 2024, blue_pen5805, the creator of anima_pencil-XL, released Illustrious_pencil-XL, a combination of Illustrious-XL and anima_pencil-XL!

illustrious_pencil-XL - v2.0.0 | Illustrious Checkpoint | Civitai

This allows natural language input, thanks to the anima_pencil-XL integration.

Illustrious_pencil-XL Workflow

Let’s update the earlier workflow to use Illustrious_pencil-XL as the base.

flowchart LR
subgraph SDXL
A(Original<br>illustrious_pencil-XL)
end
subgraph Flux.1
B(Redrawing<br>blue_pencil-flux1)
end
subgraph SDXL
C(Finishing<br>anima_pencil-XL)
end
A-->B
B-->C

For finishing, I found that anima_pencil-XL still provides better results, so I’ll stick with it for now.