Stop Memorizing Prompts: MJ/Niji is Chemistry, Not a Fill-in-the-Blank Quiz

Who Am I and Why Should You Listen (Or Not)

Okay, anyone who knows me knows this: I love chaos.

I hate using the same prompt over and over. I hate sticking to one p-code forever. Not because they don't work—they probably work fine. I just find it boring as hell.

I'm the kind of person who breaks things just to see what happens. I switch prompts. I swap p-codes. I add random words. I delete half the prompt and see if it still works.

This chaotic approach made me stumble into patterns.

Patterns that might help you troubleshoot your MJ/Niji disasters. Or maybe not. I don't know your life.

Here's the deal:

If you want a "do this, get that" tutorial → This isn't it.

If you're okay with experimental notes from someone who's stepped on a LOT of landmines → Keep reading.

So yeah. This is a field manual written by a chaotic neutral engineer.

Take what's useful. Ignore the rest. Go break some stuff and learn for yourself.

Let's go. 🫡

Part 1: Prompts Are Chemistry, Not Mad Libs

The Fatal Mistake Most People Make

They treat prompts like a template:

[Subject] + [Action] + [Scene] + [Style] = Result

So they think:

"I'll just swap 'girl' for 'boy' and 'forest' for 'beach' and get a similar vibe!"

Nope.

Prompts aren't Mad Libs. They're chemistry.

Why Chemistry?

In chemistry:

• Hydrogen? Harmless.
• Oxygen? Harmless.
• Hydrogen + Oxygen + spark? 💥 Explosion.

In MJ/Niji:

• Word A? Fine.
• Word B? Fine.
• Word A + Word B + certain context? 💥 Ban hammer or hyperrealistic nightmare.

Every word reshapes the entire "flavor" of the output.

Think of it like cooking: Sugar alone = sweet. Salt alone = salty. Sugar + salt + random spice? Could be amazing. Could be inedible garbage.

You won't know until you mix it.

Real Example: The Swimsuit Context Trap

woman in swimsuit, beach, summer vibes, candid atmosphere

woman in swimsuit, bedroom, soft lighting, intimate mood

Same swimsuit. Same woman. Different context.

The combination of swimsuit + bedroom + intimate created a "suggestive semantic field."

The system doesn't judge individual words. It judges the vibe your entire prompt creates.

Like this:

• Kitchen knife in a kitchen? Normal.
• Kitchen knife in a bedroom at 3am? Sus as hell.

Context changes everything.

The Uncomfortable Truth

What works today might break tomorrow.

Why?

• Model updates
• Training data shifts
• The AI's internal probability calculations evolve

This isn't user error. This is just how the black box works.

You can't "solve" MJ/Niji once and be done. You have to keep adapting.

Part 2: Common Newbie Mistakes (That I Also Made)

Mistake 1: Over-Using Emotional Vocabulary

The problem:

People write:

melancholy woman, depressed, lost in contemplation, feeling uncertain

The system's response:

"Cool story bro. What does she LOOK like?"

MJ/Niji doesn't understand feelings. It understands visuals.

The fix:

Translate emotions into visual cues:

• ❌ "She feels melancholy and lost"
• ✅ "downcast eyes, rain-soaked hair, grey atmosphere, hunched posture"

• ❌ "confident and powerful"
• ✅ "strong upright posture, direct gaze, dramatic lighting"

Show, don't tell. It's not a novel. It's a visual prompt.

Mistake 2: Overstacking Negative Prompts

The problem:

People panic and write:

--no young, youthful, teen, teenage, adolescent, immature, childish, 
child-like, juvenile, underage, minor, kid...

What happens:

• Diluted results
• Confused system
• Sometimes WORSE outputs

More --no ≠ better.

The fix:

Keep negative prompts short and surgical.

--no young, teen, youthful

--no [20 synonyms for young]

Focus your energy on writing a STRONG positive prompt. Use --no as a scalpel, not a sledgehammer.

Mistake 3: Ignoring Artist Name Bias

The problem:

Not all artist names are created equal. Some artist names carry strong stylistic or thematic associations that affect your output way more than you think.

Example from my own experiments:

tempting smile, curvaceous build, 
in the style of Wlop, Greg Rutkowski, Kuvshinov Ilya

[same prompt]
in the style of Wlop

Only difference: Number of artists.

Why?

Wlop's style is heavily associated with sensual female characters + dreamy romantic lighting.

So when you write tempting + curvaceous + Wlop, the system sees: Risk level: HIGH → 💥 Banned

But when you write tempting + curvaceous + [Wlop + GR + KI], the risk gets diluted by mixing in more neutral artists.

Like mixing vodka with juice. Pure vodka = strong. Vodka + juice + ice = chill.

The lesson:

Some artists are "high-risk" in the system's eyes:

• Wlop (sensual characters)
• Sakimichan (same vibe)
• Luis Royo (fantasy erotica specialist)

If you want to use them, mix them with neutral artists:

• Greg Rutkowski (epic fantasy)
• Makoto Shinkai (wholesome cinematic)
• Loish (friendly approachable)

Balance the risk.

Mistake 4: Age-Stacking Overkill

The problem:

mature adult woman, 35 years old, in her mid-thirties, 
with mature features, mature appearance, mature face, 
NOT young, NOT youthful, definitely NOT a teen...

Diminishing returns. Yes, the system has a youth bias. But drowning the prompt in repetitive constraints doesn't help.

The fix:

Use layered testing, not overkill:

Stop when it works. Don't keep adding layers.

Part 3: Surviving the Content Filter (The Tricky Part)

The Core Problem

MJ/Niji's content filter is STRICT. Sometimes it feels stricter than Google. Seriously.

And here's the frustrating part: The filter doesn't look at individual words. It looks at the overall semantic field—the "vibe" created by your entire prompt.

The "Legitimate Scene" Strategy (With Important Caveats)

I've noticed that grounding your subject in a work/professional context often helps.

Examples:

Why "Reading a Book" Can Be Risky

This sounds insane, but I've noticed: Generic "passive" activities sometimes get flagged: reading, thinking, sitting, relaxing.

The "Professional Props" Trick

Adding work-related props helps establish "this is a professional/public setting":

• microphone + stage
• stethoscope + hospital
• paintbrush + canvas
• laptop + office desk
• scientific equipment + lab

These signal: "This person is WORKING, not posing suggestively."

But Let's Be Honest...

MJ/Niji's filter is a black box. Sometimes it makes NO sense:

• Prompt A with "reading book" → Works fine
• Prompt B with "reading book" (but slightly different wording elsewhere) → Banned

There's randomness involved. You can't predict it 100%.

My General Strategy

When I'm worried a prompt might be borderline:

1. Add public/work context: in a café, bright daylight, other people visible in background
2. Use "activity" rather than "pose": ❌ "woman standing" → ✅ "woman giving presentation"
3. Keep clothing/body descriptions neutral: ❌ "tight dress, revealing" → ✅ "casual blouse, professional attire"
4. Avoid "intimate" lighting/atmosphere words: ❌ "soft warm lighting, intimate mood" → ✅ "dramatic lighting, candid atmosphere"

Part 4: P-codes Have Personalities (And Strong Opinions)

Not All P-codes Are Created Equal

Through chaotic experimentation, I've learned:

• Some p-codes love detailed, structured prompts
• Some p-codes only respond to donburi-style (short, fragmented)
• Some p-codes are camera-angle snobs (seriously)

Example: The Close-Up Diva

p-code: abcd11 + bds111 (fake example code)

• Use medium-shot? → Gives you cartoon/anime style
• Use close-up or extreme close-up? → Finally gives you semi-realistic

Why? I have no idea.

Less is More (Sometimes)

If your long, detailed prompt is producing garbage: Try deleting stuff.

What to cut:

1. Psychological descriptions: ❌ "lost in melancholic contemplation of her uncertain future" → ✅ "thoughtful expression, distant gaze"
2. Redundant atmosphere words: ❌ "warm cozy intimate comfortable soft gentle atmosphere" → ✅ "warm candid atmosphere"
3. Over-specific actions: ❌ "gracefully walking while holding a coffee cup and thinking" → ✅ "walking, holding coffee"

Donburi-Style Prompts

For certain p-codes, short and punchy wins:

rainy street, neon lights, cigarette smoke,
noir mood, trench coat, 1940s detective,
dramatic shadows

No complete sentences. No flowery descriptions. Just visual ingredients thrown together.

Like ordering food: "Rice. Beef. Egg. Sauce. Done." Not: "I would like a carefully curated culinary experience..."

Simple. Direct. Visual.

Part 5: The "Realistic" Death Trap

Words That Summon Hyperrealism

Some words are cursed incantations that trigger photo-realistic nightmares:

• ❌ realistic atmosphere
• ❌ warm and soft lighting
• ❌ natural light
• ❌ Real camera terms: f/2.8, 35mm lens, bokeh, shallow depth of field

Why Camera Jargon is Especially Dangerous

The Fix: Style Inoculation

Put this at the VERY BEGINNING of your prompt:

Semi-realism, digital illustration, pseudorealistic character art,

Think of it as a vaccine. You're telling the system upfront: "Everything I say next? Interpret it as ART. Digital painting. NOT a photograph."

Then even if you use words like detailed skin texture, the system reads it as painted detail, not photo detail.

Example Comparison

woman, detailed face, natural lighting, realistic skin

Semi-realism, digital illustration, pseudorealistic character art,
woman, detailed face, dramatic lighting, realistic skin

Also notice: natural lighting → dramatic lighting

"Natural" sounds photographic. "Dramatic" sounds artistic.

It's vibes. Vibes matter.

Part 6: Negative Prompts Are Surgery, Not Carpet Bombs

When to Use --no

✅ Use when:

• System obsessively adds unwanted elements (glasses, hats, etc.)
• Your positive prompt is clear but system ignores you

❌ Don't use when:

• Your prompt is already well-structured
• You're worried --no might backfire

The Pink Elephant Problem

If I say "DON'T think about a pink elephant," what happens?

You immediately picture a pink elephant.

Same with --no.

If you write --no photorealistic, hyperrealistic, realistic, you just made the system focus HARD on the concept of "realistic."

Sometimes this backfires. I've had prompts where adding --no realistic made things MORE realistic.

My Strategy: Surgical Precision

Keep it short. Target the exact problem.

--no eyeglasses, glasses

--no eyeglasses, glasses, spectacles, frames, optical devices, 
reading glasses, sunglasses, goggles, monocle...

More words = more chaos.

Part 7: Semantic Field Landmines

The Brutal Truth

MJ/Niji doesn't judge individual words. It judges the overall vibe your words create together.

Real Example: Artist Name Risk

From my experiments:

tempting smile, curvaceous build,
in the style of Wlop, Greg Rutkowski, Kuvshinov Ilya

[exact same]
in the style of Wlop

Why?

Wlop = sensual female characters + romantic lighting.

The system's "risk calculator":

Version A: tempting (risk +20) + curvaceous (risk +15) + mixed styles (risk +15) = TOTAL: 50% → ✅ Pass

Version B: tempting (risk +20) + curvaceous (risk +15) + Wlop (risk +30) = TOTAL: 65% → 💥 Banned

It's dilution. Mix risky elements with neutral ones.

Other Landmine Combos

1. Clothing + Location: ✅ swimsuit + beach | 💥 swimsuit + bedroom
2. Body + Pose: ✅ athletic build + running | 💥 curvaceous + lying down
3. Age + Clothing: ✅ mature woman + any outfit | 💥 young girl + ANY clothing description (instant ban, correct behavior)
4. Mood + Lighting: ✅ dramatic mood + cinematic lighting | 💥 intimate mood + soft warm lighting

The Lesson

Individual words ≠ the problem.
Word combinations = semantic field = what gets judged.

Your prompt is a recipe. Individual ingredients are harmless. Combined wrong? Food poisoning.

Part 8: Age Control is War

The Problem

Even if you write mature adult woman, in her mid-thirties, 35 years old, system gives you someone who looks 23.

Why? Training data is FLOODED with young characters. The model has a strong youth bias.

My Layered Defense

Don't use all layers at once. Build one at a time.

Part 9: Disaster Protocols

🔴 Disaster A: Hyperrealistic Hell

Symptom: Output looks like a photo. Lost all painterly beauty.

🔴 Disaster B: Unwanted Elements (Glasses, Hats, etc.)

Symptom: You didn't ask for glasses. System gives everyone glasses.

Parallel testing:

• Route A: --no eyeglasses, glasses
• Route B: Try different p-code

Test both simultaneously. Find out if it's prompt or p-code.

🔴 Disaster C: Wrong Age

Symptom: Asked for 35-year-old. Got teenager.

🔴 Disaster D: Too Cartoon/Anime

Symptom: Wanted semi-realistic. Got anime.

🔴 Disaster E: Content Ban

Symptom: Changed one word. Got banned.

It's a semantic field problem.

Solution A: Change the risky element
Solution B: Add safe context to scene
Solution C: Add third element (dilution)

Example: 💥 swimsuit + bedroom → ✅ swimsuit + bedroom + "unpacking beach bag, vacation prep"

Build a narrative that makes the combo logical and non-suggestive.

Part 10: Build Your Own Intuition

My "feeling" came from chaos, failure, experimentation. You can't download my intuition. But you can build your own.

How I Did It

1. I failed a lot - Every failure taught my brain a pattern.
2. I tested one variable at a time - Change one thing. Test. Learn which change did what.
3. I tested hypotheses in parallel - Don't test sequentially. Test multiple fixes at once.
4. I accepted randomness - Sometimes I don't know WHY it works. And that's okay.
5. I learned when to retreat - Know when to change p-code, rewrite prompt, or try again tomorrow.

What You Should Do

🎯 Step on your own landmines - Build YOUR map. My landmines aren't yours.

🎯 Experiment fearlessly - Worst case? Bad image. Generate another one.

🎯 Build feeling through repetition - After enough attempts, your brain will whisper: "This prompt feels... off. Change that word." That's intuition forming.

Part 11: Community, Collaboration, and Witchcraft

The Uncomfortable Reality

No matter how disciplined you are, randomness plays a role.

Model updates. Platform changes. Black-box probability shifts.

What works today might not work tomorrow. And vice versa.

The Real Mastery

Learning to:

• Embrace failure
• Adapt quickly
• Share knowledge
• Stay skeptical of "guru formulas"

Encourage Skepticism

Don't blindly trust ANY guide (including mine). Test everything yourself.

Ask: "Why does this work? Why doesn't that work?"

Community knowledge is built on confusion as much as success.

Share Your Experiments

Post your findings. Compare notes. Reverse-engineer successful prompts together.

Collective knowledge > individual genius.

Final Thoughts: This is a Map, Not the Territory

I'm not giving you "the solution." I'm giving you a map of where I've been.

But the terrain changes. Updates happen. Model behavior shifts.

Your journey will be different.

What I Hope You Take Away

• ✅ Prompts are chemistry, not templates
• ✅ P-codes have personalities
• ✅ Context matters more than individual words
• ✅ Artist names carry semantic weight
• ✅ Troubleshooting = controlled testing, not panic
• ✅ Sometimes it's just witchcraft
• ✅ Build your OWN intuition through failure

Most importantly:

These patterns can't be "installed." They grow through:

• Trial
• Error
• That moment when you think "wait, what if I just delete this word?"
• And it works
• And you don't know why
• But it works

That's the game.

Closing Note: On Gurus and Formulas

Be skeptical of anyone claiming to have "the answer." Including me.

Even the best pattern, overused, becomes a trap.

The model evolves. Your tactics must too.

Stay curious. Stay adaptive. Stay ready to break your own rules.

Welcome to the chaos. ❤
Good luck. May your prompts render without bans. 🎨

P.S.
If you like neat, predictable systems → MJ/Niji will frustrate you.
If you like breaking things to see what happens → You'll have fun.
Either way: Experiment. Fail. Learn. Repeat.
PuppyJun