Model vs LoRA Foundations | PixAI Mastery, Part 1

CHAPTER 01

FOUNDATIONS

— PIXAI MASTERY —

1/5 SERIES

— PART 1 OF 5 · ROOKIE —

Model vs LoRA
— The Two Building Blocks

Every image on PixAI starts with a Model — the painter, the brain that already knows how to draw. From there, you can optionally add one or more LoRAs — sketchbooks the painter consults to nudge the result toward something specific.

▸ Open PixAI

📚 PIXAI IMAGE GENERATION MASTERY · 5-PART SERIES

Part 1: Model vs LoRA Foundations ← you are here

Part 2: The PixAI Prompt Formula · Rookie

Part 3: LoRA Stacking Guide · Rookie

Part 4: AI Art Composition Guide · Advanced

Part 5: Cinematic Lighting & Depth · Master

Spend ten minutes on PixAI and you’ll see two words everywhere — Model and LoRA. You pick them in generation panel, you read about them in every community post and tutorial. And until you know what they actually are, tthe rest of the system feels like a slot machine. You click around. You try things at random. You wonder why one combination produces magic while another produces noise.

This article fixes that. By the end you’ll know what a model is, what a LoRA is, why they’re different, and the one rule that tells you when you need just one versus when you need both. Once that clicks, the rest of the series — prompt writing, stacking, composition, lighting — gets much easier to follow.

— PART ONE —

Model — The Painter

A Model is the foundation of every AI image. It’s the brain doing the painting — the part that knows how a face is structured, how light falls on fabric, how anatomy fits together, how an anime style differs from a photorealistic one. Hit “generate” on PixAI and what’s running underneath is a model reading your prompt and turning it into pixels.

The simplest way to think about it:

▸ MENTAL MODEL

A Model is “a painter who has already learned how to draw.”

Different models are different painters. They have different specialties, different aesthetics, different strengths and weaknesses:

WHAT A MODEL DECIDES

The overall aesthetic — anime, semi-realistic, 2.5D, photorealistic
Whether human anatomy holds together or quietly falls apart at the wrists
How well it handles light, shadow, and composition
The default color palette before you ask for anything (some Models lean pastel, some lean punchy)
The overall line quality — crisp, soft, sketchy

PIXAI EXAMPLES

Tsubaki.2 — newest flagship, refined details, modern anime aesthetic
Tsubaki — versatile anime base, balanced expression
Otome V2 — softer, low-saturation, romantic
Haruka V2 — strong highlights, cute-leaning faces
Hoshino V2 — sharper features, more mature look

Here’s the important part: you only get one model at a time. One painter, one canvas. You don’t stack two models together. If the painter’s style doesn’t match what you want, you don’t try to “add a second painter” on top — you either swap to a different one, or you keep this one and adjust it with LoRAs.

That’s where the second piece comes in.

— PART TWO —

LoRA — The Sketchbook

A LoRA (Low-Rank Adaptation) is a small specialist file that sits on top of a Model and pushes it in a particular direction. It’s not a Model. It’s a modifier. Without a Model underneath, a LoRA does literally nothing. Stack it on a Model and it can shift the output toward a specific style, a specific character, or a specific fix.

The mental image:

▸ MENTAL MODEL

A LoRA is “a reference sketchbook the painter consults” — or a stylistic filter they apply.

The painter (the Model) already knows how to draw. The reference book (the LoRA) shows them: draw it like this. LoRAs are small — usually tens of megabytes. They tend to do one of four jobs.

JOB 01 · STYLE

Reinforce a visual style

A “Korean illustration style” LoRA. A “watercolor” LoRA. A “ghibli-esque” LoRA. The LoRA pulls the painter’s output toward that specific look.

JOB 02 · CHARACTER

Lock a specific character

An OC LoRA, a named-anime-character LoRA. Trained on 30-100 reference images, it gets that character’s hair, eyes, signature features consistently right.

JOB 03 · DETAIL

Fix or enhance a detail

“Better hands.” “More detailed eyes.” “Cleaner line art.” These exist because base models have known weak spots, and a focused LoRA can patch them.

JOB 04 · ELEMENT

Add a specific element

“Cat ears.” “Sweater paws.” “Mecha aesthetic.” A LoRA can teach the painter a single new motif on top of what they already know.

Unlike models, you can use multiple LoRAs at once. That’s stacking — and it’s where things get interesting. Part 3 covers stacking in full. But get the foundation first.

— PART THREE —

Side by Side

To make the difference concrete, here’s how Models and LoRAs compare on the dimensions that matter:

DIMENSION

Role

Model: The base painter

LoRA: A specialist modifier

Required?

Model: Yes — every image needs one

LoRA: No — fully optional

How many at once?

Model: Exactly one

LoRA: Multiple (stacked)

File size

Model: Large — usually 2–7 GB

LoRA: Small — usually 50–200 MB

Strength control

Model: All or nothing — you use it fully or swap it out

LoRA: Adjustable from 0 to ~1.2

When you need it

Model: Always

LoRA: When the Model alone doesn’t quite get there

— PART FOUR —

The Rule for Choosing

People agonize over which LoRA to pick when the actual question is do I need one at all? Honest answer for most beginners: not yet. Here’s the rule.

RULE 01

Start with model only.

Pick a Model that already leans toward what you want. Generate a few images with no LoRAs at all. See what comes out. Most beginners are surprised — a well-chosen Model gets you most of the way there with nothing else loaded.

RULE 02

Add a LoRA only when there’s a specific gap.

Look at your output and name what’s wrong. “Hands are mangled.” “Too washed out.” “Doesn’t actually look like the character I wanted, just a generic version of her.” Each named gap is a LoRA selection criterion. Without a named gap, you’re just adding entropy.

RULE 03

Don’t ask a LoRA to fix what the model is fundamentally bad at.

If your base Model spits out grayish, washed-out output and you want vibrant punchy color, no LoRA is going to save it. Switch to a more vibrant base Model. LoRAs adjust. They don’t rebuild.

This third rule trips up almost every beginner. The instinct is to commit to whichever Model you started with and try LoRA after LoRA to fix it. It almost never works. The Model sets the shape of what’s possible. LoRAs only adjust within it.

— PART FIVE —

Working Together

Once you’ve picked a Model, LoRAs are how you fine-tune it. Here’s what a stack actually looks like.

▸ EXAMPLE STACK
BASE   Tsubaki.2 → the painter
LoRA 1   a style LoRA × 0.7 → shifts the look
LoRA 2   a detail LoRA × 0.5 → patches a known weak spot
PROMPT   describes what to draw
= a focused result

Read it like a recipe. The Model is the painter. Each LoRA is a reference page the painter flips to: “draw with this style,” “fix this detail.” The strength multiplier (× 0.7, × 0.5) controls how strictly the painter follows each one.

That’s the basic shape. Picking which LoRAs to stack, what strengths to set, how to keep them from fighting each other — that’s its own topic, and it’s exactly what Part 3 covers. For now, just hold onto the structure: one Model + zero or more LoRAs, each at its own strength.

One more piece sits alongside all this: the prompt. The prompt describes what to draw — subject, action, scene. The Model + LoRAs decide how. Writing prompts well has its own structure, and that’s Part 2.

— FAQ —

Common Questions

Can I use a LoRA without a model?

No. A LoRA on its own does nothing — it’s a modifier that needs a base model to modify. PixAI’s generator always has a model selected; the LoRA slot is optional. Think of LoRAs the way you’d think of seasoning: useless without a dish to season.

Will any LoRA work with any model?

No, and this is the single most common beginner mistake. LoRAs are trained for a specific architecture and they only work cleanly on Models built on that same architecture. PixAI shows the architecture on most Model and LoRA cards, but not always — user-uploaded Models and self-trained LoRAs sometimes leave it blank. When the architecture isn’t labeled, check the description, the example images, or the comments before loading it. If a LoRA seems to be doing absolutely nothing, architecture mismatch is the first thing to check.

If LoRAs are smaller and more flexible, why not just use lots of them?

Because they fight. Each LoRA pushes the output in some direction. A pile of LoRAs pulling against each other produces muddy, unstable, weirdly inconsistent results. Two or three is the practical sweet spot. Beyond that, you’re usually compensating for a wrong base model rather than fixing real gaps. Part 3 of this series goes deep on stacking strategy.

What’s a “trigger word” and do I need one?

Many LoRAs need specific keywords in the prompt to fully activate — a character LoRA might need the character’s name, a style LoRA might need a tag like korean illustration style. The LoRA’s model page tells you which trigger words it expects. If a LoRA seems dead, missing trigger words is the second most likely cause — right after architecture mismatch.

How do I find good models and LoRAs on PixAI?

PixAI’s Market is the central library — search by name, browse by category, sort by popularity. The fastest learning shortcut: scroll the community feed, find images you love, then check what the creator used. The metadata is on most public posts. Build your own short list of “models I trust” and “LoRAs I know what they do.” That’s where real fluency comes from.

CHAPTER 01 // END

— FOUNDATIONS COMPLETE —

— READY FOR PART 2 —

Now Pick a Model and Generate

Open the generator. Pick a Model that looks close to what you want. Run it with no LoRAs first. Look at the gap between what you wanted and what came out. That gap tells you what — if anything — to add next.

▸ Open PixAI