Reflecting on a Year of AI Image Generation: Adobe Firefly, SDXL, and Flux.1

Others #Adobe Firefly #Flux.1 #SDXL

2025-1-292025-2-17

Girl happy on her first trip to space.png (1600×1600)

Introduction

Hello, this is Easygoing.

Today, I’d like to take a moment to reflect on the past year in AI image generation.

My First Reflection Post

In my previous post, I was fortunate to receive a comment saying, "I'd love to see your older illustrations."

While I frequently experiment with new technologies, I rarely take the time to reflect on my past work.

Theme illustration for the first AI image generation journey, an anime illustration of a young woman enjoying space travel.png (672×672) — AI Image Journey

Over the past year, AI image generation has evolved dramatically, and as a user, I’ve been constantly experimenting and exploring new ways to utilize it.

Rather than focusing on AI's evolution, this post will highlight how my approach as a user has changed over time.

In writing this, I made some new discoveries myself, so I hope you find them interesting!

It All Started with a Need for Illustrations!

I started using AI image generation in November 2023, when I was writing a medical blog—quite different from what I do now.

Since my content was somewhat technical, I wanted to make it more engaging by adding illustrations.

nurse caring child at hospital.jpg (1920×1382) — Free stock material from Pixabay

At first, I used free illustration assets from sites like Pixabay.

While free images were copyright-free, I felt a bit uncomfortable using real model photos.

Around that time, AI image generation was gaining attention, so I decided to give it a try.

Trying Out Adobe Firefly

The first AI image generator I used was Adobe Firefly.

Adobe Firefly - Free Generative AI for creatives

I chose Firefly because I was already using Photoshop and was familiar with Adobe’s services. Plus, I trusted the reliability of a big-name platform.

Amazed That I Could Create Illustrations!

Generating images with Firefly was a series of surprises.

Photo of a foreign male doctor sitting in a chair in front of his desk, reviewing a paper with test results.jpg (2048×2048) — My first generated image—so realistic it shocked me

An elderly Japanese man is sitting in a hospital examination room with his back to the patient. Photo with herpes zoster skin rash on his back.jpg (2048×2048) — AI-generated images make certain depictions feel less awkward

I was amazed that I could create realistic illustrations just by typing text.

As I adjusted my prompts, the composition and expressions changed, which reminded me of the fun of photography.

Developing an Interest in AI

For about six months, I used AI-generated illustrations for my blog. But gradually, I became more fascinated by AI image generation itself than by blogging.

AI was evolving at an astonishing speed, and I felt like I was witnessing a historic moment in its development.

Since life only happens once, I decided to dive headfirst into this wave of change.

Firefly Cyberpunk, girl, crowd, electric light, realistic hologram with neon and color transitions 6.jpg (1792×2304) — Firefly struggled with anime-style illustrations

As I became more experienced with Firefly, I started noticing its weaknesses. This made me eager to explore other AI image generators.

Starting My AI Image Journey!

Wanting to document a beginner’s experience with AI image generation, I started my AI Image Journey Blog in July 2024.

first cover illustration girl in space journey — First cover illustration of AI Image Journey Blog site

Starting new blog brought many new insights. When explaining AI to others, I found myself researching topics more deeply, and it also served as a useful personal journal.

Getting Into Stable Diffusion XL

In July 2024, Stable Diffusion 3 Medium was a hot topic in the AI community.

While SD3 Medium was impressive, it seemed too complex for a beginner. So, I decided to start with Stable Diffusion XL (SDXL), which had more beginner-friendly guides available.

Stability AI, which madeStable Diffusion XL available for free.
AUTOMATIC1111, the developer behind the widely usedStable Diffusion web UI.
Zuntan03 (Zuntan), who created EasySdxlWebUi to make SDXL more accessible.

Standing on the rooftop of a skyscraper,the sprawling nightscape sparkled like an overturned jewelry box. Countless lights twink.png (1024×1024) — Capturing character emotions was difficult in Firefly

4040494876-2807929592-Without breaking the silence of the night,a black fox stood on the rooftop of a skyscraper. Its fur was as dark as the night its.png (1280×672) — Diverse anime-style illustrations

I was blown away by SDXL’s ability to generate expressive anime-style illustrations.

Anime-style art made it easier to capture emotions.

Also, unlike photorealistic images, AI-generated anime art felt more natural to me (though I imagine professional illustrators might have different opinions).

A Japanese woman in kimono eating buckwheat noodles in an elegant manner at a buckwheat noodle shop in the countryside during th.png (1024×1024) — Eating noodles surprisingly difficult to generate

At first, anime models were challenging because they required specialized tag-based prompts.

However, I discovered anima_pencil-XL by bluepen5805, which had excellent prompt adaptability. Since it allowed natural language inputs, similar to Firefly, I’ve been using it almost exclusively ever since.

High Resolution is Beautiful, But...

One of the first things AI illustration users do to improve image quality is increasing resolution.

While high resolution makes illustrations more visually appealing, it also brings out the AI's tendency toward perfection.

witch with red cloth standing In wheat fields — Flawless facial features

At first glance, perfect illustrations seem attractive, but over time, they can feel overwhelming.

4040495998-3968414304-In a certain kingdom of medieval Europe, vast wheat fields glistened in golden hues, swaying gently in the wind. Amidst this tra.png (2048×2048) — Incorporating oil painting techniques to break the perfection

Then, I came across an article by K_Kameno and started experimenting with combining oil painting techniques with anime-style illustrations by incorporating the names of Western art masters into my prompts. This helped me explore my own originality.

What Makes a Model Truly Unique?

The anima_pencil-XL model is known for being easy to use, even for beginners.

Many SDXL models recommend using quality-enhancing prompts and negative prompts.

However, the anima_pencil-XL model does not require such adjustments.

2024-08-17-1230344778-The party hall was aglow with vibrant lights, accompanied by lively music. In the center of the elegantly decorated dance floor,.png (2048×2048) — Removing "masterpiece" results in a rounder, more natural expression

20240902_020358-330318386-On the night of the masquerade ball, a sensual atmosphere pervaded the grand hall of the opulent castle. The flickering flames o.png (2048×2048) — Masquerade ball – My personal pinnacle of SDXL-generated art

So, I experimented by removing "masterpiece" and other quality-enhancing and negative prompts.

As a result, the characters’ expressions became more natural and expressive, leading me to believe that this is the true strength of the anima_pencil-XL model.

The Shock of Flux.1!

In August 2024, a groundbreaking event occurred—the release of Flux.1, an AI image generator on another level.

Realistic illustration of a cat.png (1440×1440) — Incredibly realistic textures

girl pilot in fighter cockpit — The signature round faces of the blue_pencil series remain intact

Until then, AI image generators primarily imitated real photos and illustrations.

But with Flux.1, AI could now produce results that, in some cases, surpassed actual photographs.

Flux.1 was originally designed for photorealistic rendering, but blue_pen5805 quickly released an anime adaptation, blue_pencil-flux1, which has since become my go-to model.

The High Fidelity of Flux.1 vs. SDXL's Unrefined Charm

Flux.1 delivers incredibly high-quality textures and polish.

However, because of its refined nature, it tends to produce similar-looking illustrations, limiting variation.

2024-10-01-034402_277048345144080_SDXL_Rough.png (1376×736) — Initial sketch with SDXL

2024-10-01-035649_277048345144080_SDXL_Refine1.png (3456×1840) — Final rendering with Flux.1

As I used Flux.1 more, I came to appreciate the versatility of SDXL’s less-polished models. This led me to explore combining the strengths of both.

Staying True to the Basics for Higher Accuracy

With Flux.1, I actively experimented with new technologies.

One of the most surprising discoveries was the improved CLIP-L.

Anime illustration of the future showing a woman staring at a flying car, normal CLIP-L.png (2576×2576) — Left – Standard CLIP-L, Right – Improved CLIP-L

Anime illustration of the future showing a woman staring at a flying car, improved CLIP-L.png (2576×2576) — Left – Standard CLIP-L, Right – Improved CLIP-L

Initially, I didn’t think text encoders affected image quality.

But when I tried the improved CLIP-L, the entire illustration became much clearer, which shocked me.

Anime illustration of a champagne gold night view of an amusement park Flux1 18.png (2576×2576) — Higher accuracy results in clearer illustrations.

Modern AI is still an imperfect, noise-filled analog system.

With ChatGPT’s help, I created a page to objectively evaluate image differences.

I also explored FP32 format and new text encoders, aiming to improve image quality by staying true to the fundamentals.

Letting AI Generate Freely

AI has learned from over 5 billion images.

Highly trained SDXL models possess incredible composition skills.

Flux1_Redraw_ A photo-realistic shoot from a frontal camera angle about a woman in a sleek metallic blue dress standing next to a luxurious car at night the image also shows a cityscape with blurred l_cleanup.png (2407×2407) — Difficult-to-capture angles in photography

At first, my goal was to control AI and make it generate exactly what I wanted.

However, over time, I realized that the best approach was to let AI generate freely and gently guide it toward a polished final image.

Blending AI with Traditional Digital Techniques

While AI-generated art is evolving, it still doesn’t surpass traditional digital tools.

Since the rise of digital photography in the 1990s, digital image editing has had over 30 years of development, with countless refinements along the way.

Recently, I started integrating HDR processing, a well-established digital image enhancement technique, into my AI-generated illustrations.

standard image of airplane at night — Standard AI-generated image

high quality image of airplane at night — High-precision sampler(heunpp2 beta) + HDR processing using SuperBeasts

AI models that replicate HDR processing at a human-perceptual level do not yet exist.

AI remains an incomplete tool, requiring human intervention and collaboration to fill in the gaps.

The Evolution of AI Users

This post wasn't about the progress of AI itself—it was about the progress of AI users.

Looking back at my past illustrations, I found it to be a great opportunity for reflection.

Theme illustration for the new AI image generation journey, an anime illustration of a young woman enjoying space travel.png (2576×2576) — AI Image Journey continues

I would like to thank Atlas XV and sunset for encouraging me to write this retrospective.

To everyone on their own AI image generation journey—may you create amazing work in the year ahead!

Thank you for reading!