Author's Note: If you don't want to read but just want to be gifted with a magic mad libs prompt shove this in your gpu:

The Template

<lora:LORA_NAME:WEIGHT> SUBJECT doing ACTION, SETTING, LIGHTING, CAMERA, STYLE

That's the whole spell. Six slots. Now what each one means and what happens when you get it wrong.

You're not writing a description. You're casting a spell into a prediction engine and praying it sends back a grapefruit instead of a light switch. That's the job. Most people fuck it up. Here's how to not fuck it up.

Put the subject first or it won't show up at all

The model reads left to right. The first tokens get the most juice. So when you type "cinematic highly detailed professional photography of a tagine pot" you just told the model that "cinematic highly detailed professional photography" is the most important part of your prompt. Congrats. You made a prompt about lighting equipment starring a pot that may or may not show up.

Lead with the thing you want to see. The subject. The noun. Everything else is support staff.

Be specific like you're describing a crime to 911

"A dog" gets you a dog. Sure. It also gets you every dog that ever existed compressed into one average-ass dog. "A damp golden retriever standing on a hotel balcony at 2am, one paw on the railing, city lights smeared behind it, flash photography" — now the model has something to grab onto.

The model doesn't know you. It doesn't know what you mean. It barely knows what it means. It's a foreign exchange student who just arrived and only speaks JPEG. Give it details or it fills in the gaps with whatever statistical mush floats to the top.

For the love of god stop typing "masterpiece"

You wouldn't walk into a restaurant and yell "DELICIOUS FOOD PLEASE" and expect something good. You'd get the buffet at a Comfort Inn. Same energy. "Masterpiece best quality highly detailed ultra detailed" is the prompt equivalent of pouring ketchup on everything — it doesn't make anything better, it just proves you've given up.

Swap the quality-word vomit for concrete language. "85mm f/1.4" does more work than "highly detailed professional photography." "Oil on canvas, impasto, visible brushstrokes" does more than "artistic painting." The model doesn't have a quality dial. It has a vocabulary. Use the real one.

Your negative prompt is a crime scene

Every token you shove in the negative prompt steals attention from the tokens that matter. Ten targeted negatives beat thirty desperate ones. Here's what a reasonable negative looks like:

extra fingers, mutated hands, watermark, signature

Done. Ten words. Covers the real problems.

Now here's what I see you people paste into Civitai every single day:

ugly, bad, worst, low quality, low resolution, blurry, distorted, deformed, disfigured, poor, terrible, horrible, awful, gross, nasty, amateur, beginner, unskilled, poorly drawn, poorly rendered, error, mistake, glitch, artifact, watermark, text, logo, signature, border, frame, cropped, out of frame, out of bounds, duplicate, clone, copy, mutation, mutated, extra limbs, extra arms, extra legs, extra fingers, extra hands, missing limbs, missing arms, missing legs, missing fingers, missing hands, fused limbs, malformed limbs, malformed hands, malformed fingers, long neck, cross-eyed, distorted face, distorted hands...

That's not a negative prompt. That's a panic attack. Half of these contradict each other. "Distorted" and "deformed" are the same thing to the model. You just typed it twice for no reason. The model doesn't have a "make it not suck" button. It has a direction. Tell it where to go, not every possible direction to avoid. You look like someone trying to parallel park by listing every car you don't want to hit.

Weights are seasoning, not the whole goddamn meal

(token:1.3) gives it a nudge. (token:2.0) gives it brain damage. You're not making it "more" of something — you're making the model hallucinate that token at double intensity while ignoring everything else in the prompt. Start gentle. See what comes out. Add more salt if needed. You can always season again. You can't un-salt the soup.

LoRA tags go early or they go home

The model has the attention span of a golden retriever on espresso. If you bury <lora:TagineXL:0.8> at the end of your prompt like a footnote, the model already decided what it was drawing eight tokens ago. By the time it gets to your LoRA, it's like "oh sure, tagine, I guess, whatever." Put your LoRA tag right after the subject. Give it the front row seat it deserves.

Contrast is the whole game

"Warm" means nothing. "Beautiful lighting" means nothing. "Golden hour sidelight on cold blue stone" means everything. The model lives for juxtaposition. Hot and cold. Sharp and soft. Sacred and profane. Give it tensions, contradictions, things that shouldn't coexist but somehow do. That's where the magic leaks through. A tagine pot on Mars works because the absurdity is the contrast. A tagine pot on a table is just... a tagine pot on a table. Which is fine if you're selling tagines, but you're not. You're making art. Act like it.

Steal from photographers and painters, not from the "top prompts" leaderboard

"Chiaroscuro" is a real word with a real meaning. It refers to a specific technique of strong contrasts between light and dark, developed by Caravaggio and other Renaissance painters. "Rembrandt lighting" is a specific lighting setup where one side of the face is illuminated and the other falls into shadow with a triangle of light on the cheek. These are concrete, describable things.

"Cinematic" is not a concrete thing. It's what everyone types when they've run out of ideas. It means "make it look like a movie" which means nothing because movies look like a million different things. Use the vocabulary of people who actually made images with light and hands. They spent centuries inventing words for this stuff. Borrow them.

When in doubt, describe the frame

Where is the camera? How close are we? What's behind the subject? What's sharp and what melts away? "Wide angle" and "telephoto" produce fundamentally different images, and the model knows the difference. Without camera language, the model picks whatever statistical average it has for "tagine pot" and you get the same boring mid-shot every time. Tell it where to stand and it'll give you something worth looking at.

The Prompt Mad Lib

When you're staring at the empty prompt box and your brain is doing that thing where it only outputs the sound of one hand clapping in a void — which, let's be honest, is most of us, most of the time — use this. Fill in the brackets, read it back, and you've got a working prompt. No ceremony. No tables. Just fill the blanks and cast.

The Template

<lora:LORA_NAME:WEIGHT> SUBJECT doing ACTION, SETTING, LIGHTING, CAMERA, STYLE

That's the whole spell. Six slots. Now what each one means and what happens when you get it wrong.

SUBJECT — The thing. Not "a beautiful scene." The actual noun. A tagine pot. A wet cat. A haunted espresso machine. If you can't point at it in the image, it's not your subject. It's wallpaper.

ACTION — What the subject is doing or what's happening to it. "Steaming on a clay tile." "Leaping through a ring of fire." "Existing ominously on a dark pedestal." Static subjects need action from their environment. A tagine pot sitting on a table is a product photo. A tagine pot venting steam into thin Martian atmosphere is a story.

SETTING — Where this is happening. Not "beautiful background" — what the fuck does that even mean? "A rain-soaked Tokyo alley at 3am." "The surface of Mars during a dust storm." "A grandmother's kitchen with lace curtains." The setting does half the work. Don't phone it in.

LIGHTING — How light behaves in this world. "Golden hour sidelight through dust." "Neon reflections on wet pavement." "Single candle, deep shadows." Real photographers name their light. So should you. "Good lighting" isn't a lighting setup. It's a Yelp review.

CAMERA — Where the model's eye sits. "85mm portrait lens, shallow depth of field." "Wide angle 24mm, everything sharp." "Macro, extreme close-up." This is the framing spell. Without it, the model picks whatever angle it feels like, and the model has terrible taste in angles.

STYLE — The rendering flavor. "Oil painting, thick impasto." "35mm film grain, Kodak Portra 400." "Anime cel shading, bold outlines." This is where you pick which reality the image lives in. Leave it out and you get... whatever. Probably something that looks like every other image on the front page.

Examples

The Safe One (just give me something that works, I have a deadline):

<lora:TagineXL:0.8> a tagine pot steaming on a weathered blue wooden table, scattered Moroccan spices and preserved lemons, golden hour light through a lattice window casting long shadows, 100mm macro lens f/2.8, editorial food photography

The Unhinged One (make the model question why it was born):

<lora:TagineXL:0.8> a tagine pot crash-landed on Mars with dust swirling around its lid and steam venting into thin atmosphere, red terrain stretching to a pink sky, harsh overhead sunlight with no diffusion, wide angle 24mm deep focus, National Geographic meets survival horror

The One That Shouldn't Work But Does (contrast doing the heavy lifting while you stand there looking smug):

<lora:TagineXL:0.75> a tiny tagine pot in the palm of a giant's hand, the giant is out of focus and only their weathered fingers are visible, warm firelight from below against cold moonlit mountains behind, 50mm f/1.4 shallow depth of field, Renaissance oil painting, chiaroscuro

The One Where You Forgot Everything And Just Vibe (honest, we've all been here at 2am):

<lora:TagineXL:0.8> tagine pot, underwater on a coral reef, bioluminescent creatures circling, shafts of light piercing through deep blue water, macro lens close-up, dreamlike aquatic photography

The Anti-Patterns

Things that make your prompt smell like a porta-potty at a diffusion conference. Yes, I'm talking about your prompts. Yes, all of you. I've seen the front page.

The Thesaurus Puke:

❌ masterpiece best quality highly detailed ultra detailed professional photography 8k 4k UHD cinematic stunning gorgeous amazing breathtaking award-winning photorealistic hyperrealistic

Every word here is fighting for the same tokens and none of them mean anything specific. You didn't describe an image. You described your anxiety about whether the image will be good. The model can smell your fear and it doesn't care. Use concrete language or accept your fate.

The Kitchen Sink:

❌ beautiful woman in a garden with flowers and butterflies and a fountain and a sunset and mountains in the background and birds and a path and a bench and...

The model tries to include everything and the image becomes a Where's Waldo of confusion. Nobody asked for Waldo. Pick one focal point. Build around it. Let things breathe. Not every image needs to be the entire contents of your fever dream compressed into 1024 pixels.

The Negaphobe (negative prompt terrorism, and yes, you know who you are):

❌ ugly, bad, worst, low quality, low resolution, blurry, distorted, deformed, disfigured, poor, terrible, horrible, awful, gross, nasty, amateur, beginner, unskilled, poorly drawn, poorly rendered, error, mistake, glitch, artifact, watermark, text, logo, signature, border, frame, cropped, out of frame, out of bounds, duplicate, clone, copy, mutation, mutated, extra limbs, extra arms, extra legs, extra fingers, extra hands, missing limbs, missing arms, missing legs, missing fingers, missing hands, fused limbs, malformed limbs, malformed hands, malformed fingers, long neck, cross-eyed, distorted face, distorted hands...

I see you. I see you copypasting this from some "negative prompt mega list" you found on Reddit. Half of these contradict each other. "Distorted" and "deformed" and "disfigured" all map to similar tokens — you just told the model the same thing three times and wasted two tokens doing it. The model doesn't have a "make it not suck" button. It has a steering wheel. Point it somewhere. Ten words. Maximum. I'm not negotiating on this.

The Weight Lifter (bracket abuse, and this one's for the Pony veterans):

❌ (((masterpiece))), (((best quality))), ((ultra detailed)), (photorealistic:1.5), (8k:1.4), (professional:1.3), (stunning:1.6), (((intricate details)))

These weights stack and compete. The model will fry its attention budget trying to satisfy all of them simultaneously and you'll get the visual equivalent of white noise. One or two targeted weights. The rest earns its place through concrete descriptive language. I know you saw someone on the Civitai leaderboard do this. That doesn't make it good. That makes it cargo cult prompting.

Quick-Reference Mad Lib Strips

Copy, fill, cast. Don't overthink it. That's how we got "masterpiece best quality" in the first place.

Food/Object:

<lora:___:0.8> a ___(object)___ ___(doing what)___, ___(where)___, ___(light)___, ___(lens/f-stop)___, ___(style)___

Character/Portrait:

<lora:___:0.7> ___(who)___ ___(action/pose)___, ___(environment)___, ___(light source + direction)___, ___(lens + framing)___, ___(art style or film stock)___

Landscape/Scene:

<lora:___:0.7> ___(place)___, ___(time of day + weather)___, ___(dominant colors)___, ___(camera + lens)___, ___(artistic style)___

The "I Have a LoRA and a Vibe" Special (the best prompts start this way, don't @ me):

<lora:___:0.8> ___(what the LoRA makes)___, ___(where you'd never expect to see it)___, ___(lighting that contradicts the mood)___, ___(camera that shouldn't be there)___, ___(style from a different century)___

If you get a light switch instead of a grapefruit, you probably put "masterpiece" in there again. I'm not mad, I'm disappointed. Cut that shit out.