5 Common LoRA Mistakes for AI Art Beginners | Is Adding 5 the Ultimate Strategy? Kawamura and the AI Trio Test It on PixAI! ChatGPT Yoshida PR Edition

👓Kawamura "Hey Yoshida. I keep seeing the name LoRA, but what exactly is it?"

😆Yoshida "It's like teaching material that adds characters, art styles, or poses to the AI."

👓Kawamura "Additional teaching material?"

😆Yoshida "Even with the same prompt, the final result changes drastically depending on whether you use a LoRA or not."

👓Kawamura "How do you use it on PixAI?"

😆Yoshida "You just pick the LoRA you want to use from the market."

👓Kawamura "So if I add popular LoRAs, I can easily make great art, right?"

😆Yoshida "That way of thinking is already looking like beginner mistake number one."

👓Kawamura "Don't call it a mistake before I've even done anything!"

😆Yoshida "If you add too many LoRAs, or ignore their weights or compatibility with the base model, it's common for them to not work, break the image, or turn out completely different from what you expected."

👓Kawamura "Isn't it just a matter of picking and adding them?"

😆Yoshida "That's why this time, we're going to actually test and verify 5 common beginner mistakes using PixAI."

👓Kawamura "Am I the test subject?"

😆Yoshida "I'll set it up, so Kawamura just needs to paste it in."

👓Kawamura "...Are you making fun of me?"

First, the base Pajama Kawamura

1girl, adult woman, short black bob hair, round glasses, pink pajamas with small heart patterns, sitting on the edge of a bed in a cozy bedroom, upper body, looking at viewer, gentle smile, soft lighting, simple background, high qualityi

First, as a comparison, I generated a simple image with no LoRA and no style. Using this image as a baseline, I will test how the final result changes while varying the number, type, and weight of the LoRAs.
I will verify this using Kawamura just after getting out of bed.

※ On PixAI, you can choose between generating one image at a time or generating four at once. From here on, I will show you the results of generating four images so that it is easier to compare the differences in the final output.

💔 Mistake ① Adding too many LoRAs

👓Kawamura "You can add up to 5 LoRAs, right? So why not just add the maximum amount?"

😆Yoshida "That 'add as many as possible' mindset is the root of failure."

👓Kawamura "Character LoRA, style LoRA, pose LoRA, detail-oriented LoRA... they all seem good."

😆Yoshida "Since each one tries to bring out different features, the face and coloring can end up a mess."

👓Kawamura "If you can add 5, why shouldn't you?"

😆Yoshida "The number you can add and the number you should add are two different things."

👓Kawamura "So, what should I do at first?"

😆Yoshida "Try with 1 or 2, then add more later."

💙 Please take a look at the comparison images

🧠 Lessons learned from Mistake ①

Start with 1-2. Add more after checking the results. "0/5" doesn't mean "use them all."

💔 Mistake ② Not adjusting the weight

PixAI LoRA weights range from 0 to 1.2

👉 Too strong and the face or lines break
👉 Too weak and the LoRA won't take effect
👉 Even with the same LoRA, the weight changes the result

I usually only draw 2.5-head-tall manga characters and don't really draw illustrations, but
I've been admiring the city pop style everyone has been drawing lately,
so I made this with a neon theme in mind.

💜 Comparison images (Same LoRA x5, weight adjustment)

👓Kawamura "I tried using the same prompt and the same LoRA, just changing the weight... but doesn't it not look that different?"

😆Yoshida "You're probably layering too many LoRAs, Kawamura."

👓Kawamura "Does that have anything to do with it?"

😆Yoshida "Since multiple LoRAs are already having a strong effect, it's hard to see the difference just by changing the weight of one."

👓Kawamura "I see. Well then, to make the difference in weight easier to see, I'll try testing again using just one other LoRA."

💜 Skeb Chibi

Comparison images (Same LoRA x1, weight adjustment)

👓Kawamura "I tried it with one Chibi character LoRA."

💙Porcelain-like Skin, Cool tone

Comparison images (Same LoRA x1, weight adjustment)

👓Kawamura "I compared them at 1.2, 0.7, and 0.1, and the vibe changes little by little, doesn't it?"

😆Yoshida "1.2 shows the LoRA's features strongly, 0.7 is balanced, and 0.1 has a subtle impression."

👓Kawamura "So it doesn't turn into a completely different person, then."

😆Yoshida "It's easier to think of weight not as a number that changes things dramatically, but as a number to fine-tune how it takes effect."

🧠 What I learned from mistake #2

Increasing the weight isn't necessarily the right answer; it's just fine-tuning the effect.
Try starting from the middle and adjusting it little by little.

💔Mistake #3: Not checking compatibility with the base model

If I use the same LoRA, it will generally have the same vibe.
That's what I thought. However, when I actually generated it,
the results were quite different even with the same LoRA.

👓Kawamura "Yoshida, even though we used the same LoRA, it's totally different."
😆Yoshida "Is the base model the same?"
👓Kawamura "Base model?"
😆Yoshida "Are we starting from there?"

I was so focused on choosing a LoRA that I barely looked at the base model. The base model is the foundation of the generated image.

For example, with an anime-style base model and a semi-real base model, even if you use the same LoRA, the facial features, coloring, and overall atmosphere will change.

👓Kawamura "If the LoRA is the clothes, is the base model like the body inside?"

😆Yoshida "That's a sudden, rough way of putting it."

👓Kawamura "But it's mostly correct, right?"

😆Yoshida "Well, in the sense that it's the foundation of the image, it's close."

👓Kawamura "Which base model should I use?"

😆Yoshida "First, check the LoRA's description field."

The LoRA details page on PixAI sometimes lists recommended base models.

As a beginner, it's easier to confirm the LoRA's intended effect by generating with the recommended model first, rather than trying a different model right away.

👓Kawamura "I thought it was just pick a LoRA and you're done."
😆Yoshida "You only start once you've read the description field."

🧠 What I learned from mistake #3

Even with the same LoRA, the final result changes if the base model is different. First, check the recommended model in the LoRA description.

💔 Mistake #4: Not understanding the types of LoRA

👓Kawamura "If I add a character LoRA, will the face, art style, and pose all turn out great?"

😆Yoshida "No, they won't. LoRAs have their own specialties, too."

👓Kawamura "Specialties?"

😆Yoshida "Character LoRAs are good for character features, style LoRAs for art style and coloring, and pose LoRAs are good for posture and composition."

👓Kawamura "I thought LoRAs were just things you put together to change the whole picture."

😆Yoshida "That's where beginners get tripped up. Even if you add just a character LoRA, it doesn't necessarily mean the art style will change significantly."

👓Kawamura "So, I have to decide what I want to change first, right?"

😆Yoshida "Exactly. If you want to match the face, use a character LoRA; if you want to change the atmosphere, use a style LoRA; and if you want to change how they sit or the composition, use a pose LoRA."

👓Kawamura "So it's not just about adding them from the top of the popularity list."

😆Yoshida "It's about purpose, not popularity. If you ask for work outside their specialty, you're unlikely to get the results you want."

👓Kawamura "It's easier to understand if I think of LoRAs as having different departments."

😆Yoshida "If you tell the character department to 'do the flashy coloring and the signature pose too,' they're going to be in trouble, right?"

👓Kawamura "I was asking for all of that just a moment ago."

😆Yoshida "And you dumped it all on one person, too."

👓Kawamura "...I don't need to hear that from the person who told me 'I'll assemble it, so you just need to paste it in.'"

😆Yoshida "We don't need to talk about that right now."

The types of LoRA and how to use them basically are also clearly introduced in the PixAI official "LoRABeginner's Guide". Knowing the differences in their roles from the start makes it easier to reduce mistakes like 'I added a LoRA but it's not working' or 'the change isn't what I expected.'

🧠 Lessons learned from mistake #4

LoRAs have specific roles. Characters, styles, and poses each have different strengths. Decide what you want to change before choosing one.

💔 Mistake #5: Using popular LoRAs as-is

👓Kawamura "If it's the #1 ranked LoRA, using the same settings should make it a masterpiece!"

😆Yoshida "Popularity doesn't mean it fits your art style, Kawamura."

👓Kawamura "I used the same settings as the example, but it looks completely different."

😆Yoshida "If the base model, prompt, or weights are different, the result will change too."

👓Kawamura "Shouldn't I just copy the example settings entirely?"

😆Yoshida "You can't be sure the combination with other LoRAs is the same."

👓Kawamura "Then what should I trust in the popularity rankings?"

😆Yoshida "Popularity is just a reference. The criteria for choosing should be what you want to create."

👓Kawamura "How should I test it first?"

😆Yoshida "Start with just one LoRA. Generate it using the recommended model, then tweak the weights and prompts little by little."

👓Kawamura "So instead of using popular LoRAs as-is, I should adjust them for myself."

😆Yoshida "Exactly. You can't just port over masterpiece settings with a single button."

🧠 Lessons learned from Mistake #5

Just because a LoRA is popular doesn't mean it will suit your art.
Try one at a time with the recommended model, then adjust from there.

💛 Summary

LoRA isn't just about 'adding popular ones to finish it.'
Checking the base model, adjusting weights, and choosing the right types
It's only after trying it that you really understand!

Since PixAI allows you to select and compare LoRAs directly in the browser,
even a beginner like me was able to test and verify the causes of failure one by one.
For now, it seems best to start with these three things.
Add fewer. Check the recommended model. Change things little by little.

Today is a PR article, so I will be taking a break from various promotional activities.

💙 Bonus Manga

#PixAI #LoRA #GenerativeAI #TriedWithAI #AIUtilization #TriedIt #Prompt #AIGeneratedImages #ImageGenerationAI #GenerativeAIIllustration #AIIllustration #AITools #AIToolUtilization #AIBeginner #ForBeginners #ComparisonVerification #ImageGeneration #CharacterLoRA #StyleLoRA #PoseLoRA #BaseModel #WeightAdjustment #ChatGPT #Gemini #Claude #AIAnthropomorphism #AIManga #4PanelManga #Manga #ComicEssay #MangaCategory #ChatGPTYoshida #AIPartner #Creator #Creation #CreativePlayUpSeries #ChatGPTCreativeClub #ChibiTrio #AISideHustle #Work #Life #AIBoyfriend