5 Common Mistakes LoRA Beginners Make | Verifying Common Causes and Improvements on PixAI Using Sora-chan Images! | Generative AI | AI Illustration |

*This article contains a promotion for PixAI.

Hello, I'm Nenkoro🌷

I previously created tons of "soccer cheering characters by country" on
PixAI, but...

The thing I struggled with the most back then was
LoRA.

That's what I thought. At first.

......It did not.😇

Moreover, when I created several characters
with the same LoRA settings,

Their faces were so similar that I wondered that.

I thought, "Am I the only one failing?" but
I understand now.
Most of the causes were "how to use it".

So this time, after properly researching LoRA,
I've organized the "mistakes beginners often make"
into 5 points,
and I'll introduce them along with improvement methods you can do within PixAI.

Here is the base for the verification👇

I'll use the standing illustration of my signature character, Sora-chan,
created only with the model Tsubaki.2
(without LoRA) as a standard,
and verify by adding LoRA to it.

By keeping the conditions consistent,
it should be easier to see
"what changed with LoRA"😊

There must be many people
stumbling over the same things👇

Mistake 1 | Adding too many LoRAs causes the image to collapse

This is what beginners do first.

Finding popular LoRAs and adding them all🤣

During my last verification, I started by
adding character LoRAs, style LoRAs, and so on...
Before I knew it, I had stacked 4 or 5 of them.

As for what happened as a result,

• The face became unstable (a different person every time it was generated)

• The coloring mixed together, losing any sense of unity

• The character itself fell apart

In Sora-chan's case👇

She's small, cute, and sparkly!
This is nice too, but it feels a bit mismatched!

The reason this happens is because
the LoRAs influence each other.

It's like they are each insisting on their own "art style"
and end up fighting.

Improvement method: Start with only 1 or 2

This is the workflow I settled on👇

• First, generate with only one added

• After seeing the result, add one more if necessary

• Thoroughly "check before adding"

Just by narrowing down the number,
the stability of the face was completely different✨

Mistake #2 | Not adjusting the weight

I did this at first, too.

Using the weight at the default
or fixed at 1 to 1.2😇

LoRA has a strength setting called "weight"
that determines how much it affects the image,
but beginners often don't touch this
and end up worrying, "Why does it look so distorted?"

The PixAI default is0.7, but when I raised it to 1 thinking,
"I'll make it stronger," the image fell apart perfectly😇

Conversely, when I lowered it to 0.3,
it was too faint,
resulting in a "Is the LoRA even working?" state.

Improvement: Use the default 0.7 as a baseline and adjust it little by little.

In the range I tested,

• 1.0 → Distorted (LoRA's influence is too strong)

• 0.7 (default) → Stable. Start here as a baseline.

• 0.5–0.6 → A sweet spot that is less prone to breaking and allows for adjustments.

• 0.3 → The effect is too weak.

The default 0.7 is
actually very well thought out, isn't it?

The trick is to first generate at 0.7,
if it breaks, go to 0.5–0.6,
and if it's not enough, raise it slightly.
Adjusting in 0.1 increments is the golden rule💡

If you change it all at once,
you won't know what the cause is,
so the point is to do it little by little.

Mistake #3 | Not checking compatibility with the base model

When I thought, "The LoRA isn't working!"
it was usually this.

Even if you only change the LoRA,
if the compatibility with the base model is poor,
you won't get the ideal image.

For example, the above was generated with the recommended model, but
if you change the model to a Korean-style one likeSerinthe result
will be completely different even with the same LoRA.

Improvement method: Try the recommended model first

On PixAI, compatible models are often listed on the LoRA page,
so
trying the recommended model first is the shortcut.

Once you know the "correct state" with the recommended combinations,
changing to a model you prefer
makes it easier to judge whether it's working or not.

Mistake 4 | Not understanding the types of LoRA

Embarrassingly, at first I

I thought that 😇

Actually, LoRA has different roles👇

• Character LoRA: Reproduces the appearance of a specific character

• Style LoRA: Changes the art style or drawing style

• Pose LoRA: Stabilizes the composition or pose

Character LoRAs are strictly for "character duties."
If you want to change the art style, use a style LoRA; if you want to determine the composition, use a pose-based one.

Choosing according to your goal is the correct approach.

Deciding "what you want to change" before searching for a LoRA will keep you from getting lost.

Mistake 5 | Using popular LoRAs as they are

I used a top-ranked LoRA, but it didn't work out—

This is also a common occurrence.

Just because it's popular doesn't mean it's right for you. right?

• If the model or prompt is different, the results will also change.

• Even if you copy the same settings as the example, you aren't guaranteed a perfect reproduction.

• Sometimes the direction is just different from the image you want to create in the first place.

Also, this is from my own experience, but
when creating multiple characters
if you reuse the same LoRA settings,
everyone ends up with the same face🤣

It's the classic "Are you all siblings?" incident (lol)

LoRA's influence is that strong, so
if you want to create differences between characters,
it's recommended to review not just the prompts for hairstyles and expressions,
but also the LoRA combinations and strengths
on a per-character basis.

Improvement method: Adjust little by little in your own environment

• View examples as "references"

• Readjust the weight with your own model and prompts

• Find a LoRA that fits your purpose (character reproduction? art style?)

By the way, when searching for LoRA on PixAI,
if you plan to use it for articles or projects,
don't forget to check the "Commercial Use Allowed" filter as well✅

I'll leave a summary at the end!

Frequently Asked Questions (FAQ)

Q. How many LoRAs can I include?

A. "The bare minimum" is more important than the upper limit.
I recommend trying with just 1-2 first,
and only adding more if you feel something is missing.

Q. What weight should I start with?

A. Starting with the default 0.7 is fine.
If it breaks, lower it to 0.5-0.6,
if the effect is too weak, raise it slightly,
try adjusting in increments of 0.1.

Q. What should I check when LoRA isn't working at all?

A. First, check the compatibility with the base model, then check the weight.
It's easier to isolate the cause if you revert to the recommended model and test it.
Switching back to the recommended model makes it easier to identify the cause.

Summary | Start with "fewer and simpler" LoRAs

As a beginner,

• Don't add too many(start with 1-2)

• Adjust weights little by little(start around 0.7)

• Check compatibility with the model(start with the recommended model)

• Understand the types and roles

• Choose what suits you, not just what's popular

Just by keeping these 5 things in mind,
the stability of your generations will change completely.

LoRA is convenient, but
the results vary greatly depending on how you use it.

In that regard, PixAI is a great environment for beginners to "try things out little by little"
since you can complete both LoRA additions and weight adjustments entirely within the browser☺️
environment.

For those wondering what LoRA is in the first place,
the explanation on the official PixAI blog

is easy to understand,

so I recommend
reading that first.

I hope you enjoy the process, including the failures,
and create
your own unique piece of art🎨

You can play with free credits!

Well then,
thank you
for reading until the end!

This was Nenkoro!

#PR #PixAI #LetsPlayWithAI #TriedItOut #AIUtilization #ImageGenerationAI #LoRA #AIIllustrationBeginner #CharacterDesign #GenerativeAI #AIUtilizationTips #NoteArticle #Illustration #ArticleCreation #LearningRecord #InsightsAndLearning #AIUtilizationForBeginners #ImageGenerationAIUtilization