The shot starts before you generate
Most people open the AI before they know what they actually want from it. Result: dozens of attempts, burned credits and a feed where no two photos belong together.
Generating isn't “hit the button and pray.” It's a chain of decisions you make before your first prompt even goes in.
Beginners almost always start the same way.
They open the AI, type “beautiful girl” and crank out a hundred images in one evening.
What they really get is a hundred versions of the same average shot: plastic skin, a warped body, extra fingers and a face that changes from post to post.
All of it goes straight to the feed, and a month later the AI, the algorithm and the audience take the blame. But the mistake happened at step one: the shot never had a brief.
The AI is great at executing a brief. It just can't come up with one for you. No brief, and it fills in the most average thing it's ever seen. And average hooks nobody.
“I want a pretty photo of a girl by the sea.”
What you get: sunset, perfect hair, a dress blowing in the wind, a gaze into the distance. Pretty. And exactly like a thousand others.
“I need the first shot of a ‘beach weekend’ set. She's walking down a wooden boardwalk toward the beach, shot by a friend on a phone mid-stride, from 2–3 meters away, slightly behind and to the side. Camera at chest height, horizon a little tilted, like it was taken without stopping. She's half-turned, looking back over her shoulder and laughing, hair messed up by the wind. About an hour before sunset, the sun low and off to the side, warm light and long shadows on the boards. She's wearing a light linen dress and flip-flops, a beach bag in her hand. In the background: sand, a slice of sea and a couple of people far away, out of focus. Slight motion blur at the edges, ordinary phone-camera quality, zero studio sharpness.”
Now the brief has everything that later goes into the prompt: who's shooting and on what (a friend, a phone), distance and angle (2–3 m, behind and to the side, chest height, tilted horizon), action (walking, looking back), emotion (laughing), time and light (evening, low sun from the side), outfit and props, background and “real-life” imperfections. The AI has no gaps left to fill with a template.
Before you write the prompt, answer one question in one sentence: why does this shot exist? No answer means it's too early to generate.
Three questions before your first prompt
Where will this shot live?
Profile pic, part of a set, a feed video, a story photo. That decides the format, how much detail it carries and how “accidental” it should look.
What does it need to do?
Stop the scroll. Show her personality. Move the story forward. Add some breathing room between strong shots. One shot, one job.
Will viewers believe it's real?
Who was holding the camera? Where's the light coming from? Why is that in the background? If there's no logical answer, the viewer will sense a fake before they can tell you why.
Three core formulas
Every photo and video of an AI model is built from three simple formulas. The tools change. The formulas don't.
Photo
The AI takes the face from a photo and builds the scene from text.
Video: Image to Video (I2V)
A finished photo comes to life. The look, outfit and location come from the photo, the motion comes from text.
Video: Motion Control
Your model copies the moves from someone else's clip: a dance, a trend, a gesture, a walk. The reference gives the sample, the prompt gives exact instructions on how to perform it.
The formulas form a chain. Video is made from photos, photos are made from the face. A mistake at an early step doesn't disappear: it moves down the line and gets more expensive.
The tool names below are examples, not a forever recommendation. For photos people often use Nano Banana Pro, for video Seedance and Kling. The models get updated every few months, but each formula's inputs and outputs stay the same. So learn the formula, not one app's buttons.
Reference face: where it all starts
The “face + prompt” formula is exactly as good as your reference face. No prompt can fix a bad one.
The reference face is a clean, well-lit shot of the face that keeps your model recognizable. It goes into every photo generation. And it's what you compare every result against.
No face yet? First generate a batch of portraits from a text description of the look and pick the one you'll run with. That one becomes your reference face. Face mixing and locking in the look (with LoRA, for example) are a separate, deeper topic. To get started, one good portrait is enough.
What a good reference face looks like
Straight-on, looking into the camera, head not tilted
Even, soft light, no hard shadows across half the face
No filters, beauty retouching or heavy editing
Face fully visible: no hair, hands, glasses or phone covering it
Calm expression or a slight smile, not a grimace
Sharp, good resolution, the face fills a good chunk of the frame
Simple clothes, minimal jewelry. Whatever's in the reference face photo, the AI will drag into every shot
Up-close front-camera selfie, evening, a warm lamp off to the side, big smile, a strand of hair on her cheek.
A wide-angle camera this close distorts proportions: bigger nose, smaller ears. The shadow hides half the face, the smile changes the shape of the cheeks and eyes. The AI will carry all of that into every shot as “your model's features.”
Head-and-shoulders portrait, shot from 1.5 meters, daylight from a window in front, neutral face, hair pushed back behind the shoulders.
You see the real proportions and every feature. A face like this drops in just as well at a café, on a beach or in a mirror selfie.
Many generators take several references at once. Two or three extra angles of the same face (3/4, profile) help hold it in tricky poses. The main rule: use the exact same reference set in every generation, not “whatever's lying around.”
Drift: how your model “wanders off”
In “AI OFM Fundamentals” I called an unstable look one of the biggest problems with AI content. Here's what it looks like in practice.
Matches the reference face.
The chin has changed shape a little: narrower, sharper. Compared to shot 2: “looks the same.”
Different eyebrow shape. Compared to shot 5: “fine.”
Put it next to shot 1: it's a different person. Every step was small, the total is huge.
I see this with beginners all the time: they run a model for two months, and her own followers no longer recognize her in the new shots. Nobody caught the moment she “wandered off,” because every new shot was only compared to the one before it.
Comparing to the previous shot hides drift. Comparing to the reference face exposes it instantly. Another common cause: building a new scene from the last generation instead of the reference face. The AI takes every mistake in the previous shot as truth. Every new scene starts from the reference face.
What should change, and what never does
Should change
Hairstyle, outfit, location, light, mood, angle. Without that, your feed turns into the same shot in different versions.
Never changes
Facial features, overall look, body proportions, signature marks: a mole, a scar, a piercing, a quirk in her smile.
The mark that saves you
Real people always have some small thing you recognize them by. Give your model one up front and keep it in every shot.
It pays off twice: viewers recognize her faster, and you notice faster when the model starts to “slip.” Mark missing? The shot gets redone.
Photo: face + prompt
The AI takes the face from the photo. Everything else comes from text. So the quality of the shot depends almost entirely on how the prompt is written.
“Big eyes, full lips, sharp cheekbones” in the text fights the reference. The AI tries to please both the photo and the text and ends up averaging them. The face starts to drift.
One line is enough: “use the face from the reference.” Give the rest of the prompt to the scene.
What goes into a scene prompt
Every shot breaks down into the same layers. Skip a layer and the AI fills it in on its own. Almost always worse than you need.
How it's shot
Selfie, mirror, shot by a friend, camera on the table. It's the first thing the eye reads.
Angle
Camera height, tilt, distance, how tight the framing is, where the model sits in the frame. Without it, the AI plants her dead center like a passport photo.
Where
Not “a café” but a specific café: a table by the window, a sugar bowl, someone else's jacket on the next chair.
Light
One clear source and which side it's on. Window on the left, lamp behind, overcast street. Light “from nowhere” gives the generation away instantly.
Pose and action
Fixing her hair, holding a glass, looking at her phone. Action turns a shot into a moment instead of a pose.
Emotion
Sleepy, laughing, lost in thought, someone just called her name. One emotion, not four.
Outfit
Cut, color, fabric, how it fits. “Stylish clothes” is not a description.
Realism
Real skin, slight phone-camera noise, an ordinary everyday background. Write what the result should be: “visible pores,” “blurry bg.”
Example breakdown
Girl in a café, beautiful, stylish, good lighting, high quality.
There's not a single decision here. “Beautiful,” “stylish,” “good,” “high” are ratings, not details. The AI will ignore them or hand you glossy ad polish.
Shot by a friend on a phone at a table by the window. Overcast daylight falling from the left. On the table: a half-finished cappuccino and a phone face down. Chunky grey sweater, hair loosely tied up, a couple of strands falling out. Looking just past the camera, half-smile, like someone off-camera just said something.
Every word is a decision. Not a single rating. You've got who shot it, where the light comes from, what's around, what she's wearing and what's happening.
The same shot as a finished prompt
Prompts are usually written in English anyway, since most generators understand it best. Same structure, layer by layer.
Use the woman from the reference image as the only subject. Keep her facial features exactly as in the reference. Shot: casual smartphone photo taken by a friend sitting across the table, vertical 3:4. Camera: eye level, about 1.2 m away, slight tilt to the right, framed from the waist up, subject slightly left of center. Location: small European café, table by the window, half-finished cappuccino, a black iPhone 15 Pro lying face down on the table. Light: soft overcast daylight from the window on the left, neutral color, no studio light. Pose and action: sitting sideways on the chair, one elbow on the table, looking just past the camera. Emotion: relaxed half-smile, as if someone off-camera just said something funny. Outfit: oversized light grey knit sweater, straight blue jeans. Realism: natural skin texture with visible pores, faint phone-camera noise, unretouched look. Hands: one hand around the cup, the other resting relaxed on the table. Background: the street behind the window softly out of focus.
Write what you want, not what you don't. Instead of “no plastic skin,” write “visible pores”; instead of “no text on signs,” write “street softly out of focus.” The AI follows a description more accurately, and the banned word sometimes sneaks into the shot.
The swap rule: replace every “beautiful,” “high-quality” and “stylish” in your prompt with the concrete detail that makes the shot that way. Not “cozy” but “a throw blanket on the armrest and a lamp behind her.”
Overload is the flip side
You need details, just not all of them at once. When a prompt has twenty objects, four light sources and three actions, the AI starts arguing with itself. Result: extra hands, a stranger's face, mush in the background.
The polish loop
A good shot rarely lands on the first try. But running the same prompt twenty times gets you nowhere either. What works is a short loop:
A reference is inspo, not a template
You find someone else's shot that hooks you. The biggest mistake is trying to copy the whole thing. The right move: figure out what exactly makes it work and take only that.
Breaking one down
A girl sitting on a windowsill in an oversized shirt, knees pulled up, the city at dusk outside the window, the room dark, her face lit only by her phone screen.
The contrast between a dark room and one cold light source. Solitude. A pose that's “for herself,” not for the camera.
Your model on the edge of the bed in her own hoodie, legs under the blanket. Dark, the only light is the laptop on her knees. Outside the window, no city, just a couple of windows in the building next door. The angle is slightly from the side, like her roommate took the shot.
Same atmosphere. Different shot. A viewer who's seen the original will feel a familiar mood but won't spot a copy.
A 1:1 copy of a popular shot works against you twice. Anyone who's seen the original stops believing in your model. And anyone who hasn't just gets one more lookalike picture in their feed.
Reference bank
Don't go hunting for inspo the moment you need to generate. Collect it ahead of time and sort it by purpose, not by “what looks nice.” For video, keep a separate folder of clips with moves you want to copy.
Sets, not one-offs
A single shot shows a picture. A set shows a life. Sets are what turn a pile of generations into someone people want to follow.
Real people don't post one photo from one place. They post a day, a trip, a night out with friends. Think in episodes, not shots.
Example: one episode, “Sunday”
Morning
Bed, crumpled pillow, phone in hand, light through a gap in the curtains.
Coffee
Just hands and a mug on the windowsill. No face, and that's fine.
Getting ready
Hallway mirror selfie in the outfit she's heading out in.
Outside
Shot from the side, walking, not looking at the camera. Same outfit as in the hallway. A good candidate to bring to life as a video.
Evening
Tired, at home, a blanket, a show on in the background. Same hairstyle, just a bit messier.
Sets break on the small stuff. If she's in white sneakers in the hallway and black ones outside, the episode falls apart. Before generating each shot in a set, check the outfit, jewelry, hairstyle and time of day.
The set card
Fill it in quickly before you start and keep it in front of you while you generate. Copy the shared lines from the card into every shot's prompt word for word. That way outfits and details don't jump around.
Sunday at home + a walk
grey hoodie, light jeans, white sneakers
thin chain, ring on her left hand
down in the morning → low ponytail by day
morning: window, day: overcast, evening: lamp
mole by her left eye (check in every shot)
“Faceless” shots count too
Hands with a mug, legs on a blanket, the view from a window, a bag on a chair. Shots like these are cheaper, almost never break and make the feed feel alive. Real people post their surroundings all the time, not just themselves.
Where the image breaks
Every AI has zones where it keeps messing up. Know them up front and you won't waste an hour on something that won't work.
Hands
Especially holding something or with fingers interlaced. An extra finger is a classic find.
Jewelry
Mismatched earrings, a chain melting into the skin, a ring that “moves” to another finger.
Text
Signs, labels and T-shirt prints turn into a mush of half-letters.
Reflections
The mirror shows a different pose, a different phone or a different person altogether.
Small objects
Forks, pens, earbuds and shoelaces melt into each other.
Unusual angles
Low angle, high angle, a heavy tilt: the AI knows these worse, and proportions slide.
If a scene runs into the tool's weak spot, don't fight the AI for twenty attempts. Change the scene.
Can't get the hands right with a wine glass? Put the glass on the table. Sweater print is mush? Lose the print. Mirror lying? Shoot from the side.
The three-second test
A viewer gives a shot a couple of seconds. Check it the same way: open it on your phone, look for three seconds, close it. What stuck first?
Her eyes, her smile, the mood. The shot works.
“Weird hand,” “something's off with the earring,” “the sign is warped.” That's exactly what the viewer will remember too. Redo the shot.
Too perfect is a mistake too
By default, AI pulls toward “quality”: smooth skin, even light, everything sharp. A real phone doesn't shoot like that.
If the shot is done but looks too clean, you can add light noise and that phone-camera character for free: FBL APPS → PHONE LOOK.
Video: Image to Video
Finished photo + motion prompt = a short clip. The look, outfit, location and light are already in the photo. The prompt only handles what changes over time.
The main rule: stable shot first, motion second. A weak photo won't become a strong video. It becomes a weak video where the mistakes move too. That extra finger in the photo will start wiggling in the clip.
The photo sets the clip's format. Need a vertical video for the feed? Generate a vertical photo from the start. Stretching and cropping later costs you quality and composition. You can check for free whether her face will end up under the platform's caption and buttons: FBL APPS → SAFEZONE.
Describe the motion, not the picture
A beautiful girl in a grey hoodie sitting on a bed in a cozy room with soft light.
The AI can already see all of that. There's no motion in the text, so it'll make some up. And retelling her looks sometimes makes it “redraw” the face mid-clip.
Looks up from her phone, looks into the camera, laughs softly and turns away. Slow movements. Handheld camera, slight natural shake, no zoom. The background stays still.
What she does, in what order, at what pace, what the camera does and what has to stay put. Nothing left for the AI to guess.
What goes into a motion prompt
Action
What she does and in what order. One or two actions, not five.
Pace
Slow, smooth, natural. Without a set pace, movements often come out jerky.
Camera
Locked in place, in a friend's hand or in her own outstretched hand. Leave it out and the camera starts “flying around.”
What stays still
Background, furniture, objects. Otherwise the walls “breathe” and the cup crawls across the table.
What to keep
Face, outfit and hairstyle unchanged until the very last second.
How to write negatives
Seedance has no separate field for negatives. Instead of “no melting face,” write “her face stays sharp and unchanged.” Describe the result you want.
@Image1 is the first frame. She is sitting on the bed holding her phone. She slowly lifts her eyes from the screen, looks into the camera, laughs softly and turns her head away. Natural breathing, small shoulder movement while laughing. All movements are slow and smooth. Camera: handheld phone held by a friend, very slight natural shake, no zoom, no pan. The background and furniture stay completely still. Her face stays sharp and unchanged, eyes keep a natural relaxed shape, hands keep a natural shape. Hair and clothing stay exactly as in @Image1 until the last frame.
The motion ladder
Motion gets harder step by step. Don't skip ahead: until step 2 works every time, step 4 can wait.
Length is a step too. Learn on 5 seconds: the longer the clip, the more the face changes by the end.
One idea, two approaches
She walks into the room, takes off her jacket, throws it on the bed, sits down, picks up her phone and laughs.
Five actions, a pose change, two objects to handle. Almost guaranteed: the jacket changes style, the hands break, and by the end of the clip it's a different face.
She's already sitting on the bed with her phone. Looks up from the screen, looks into the camera, laughs and turns away.
One pose, one action, one emotion. Same idea (a “real moment at home”) with half the risk.
Video: Motion Control
You don't need a model that's “just alive.” You need a specific move: a trending dance, a gesture, a walk. The reference shows the move, the prompt tells the AI exactly how to transfer it. Without a strong prompt, even a perfect reference gives you a sloppy result.
Motion Control comes in two flavors. In Kling it's a separate mode: the frame and the clip go into their own fields, the AI takes the motion and facial expressions from the clip, and the prompt covers whatever is in neither: the camera, the background, small details. In Seedance the reference clip is attached to a regular prompt, and the prompt carries more weight: it's what tells the AI which file does what. Below we break down Seedance.
The start frame is a photo of your model that the clip opens on. The reference is a video of a real person doing the move you want. In Seedance both files are uploaded together, and the prompt points to them with tags: @Image1 is the frame with your model, @Video1 is the video with the move. The tag format depends on the platform: some use @Image1, some [Image 1]. Check the hint in the prompt field.
Who handles what
Write a one-line prompt like “repeat the dance from @Video1” and the AI transfers the motion “more or less.” Hands smear, fingers stick together, expressions go random, the range of motion keeps growing and shrinking, the camera starts wandering.
The reference is the sample. The prompt is the instructions for performing it. You only get decent motion when both are doing their job.
What to spell out in the prompt
Who and from where
@Image1: who's in the shot. @Video1: where the motion comes from. One line at the start of the prompt.
Style of motion
Not “what and when” (the order and timing are already in @Video1) but “how”: smooth or sharp, soft or crisp arms, where the emphasis goes: hips, shoulders, head.
Arms and body
What the arms, hands, shoulders, hips and head do. The more precisely you describe the hands, the less finger “mush” you get.
Expression and gaze
Where she looks, when she smiles, when she looks away. Another person's facial expressions are the hardest thing to transfer, so spell them out separately.
Pace and range
On the beat, smooth, no speed-ups. How big the moves are: same as the reference or softer.
Camera
Completely still: no pans, tilts, push-ins, sway or horizon drift. Don't lock it down and the camera starts copying the reference's camera operator or “wandering” on its own.
What to keep
Face, body proportions, hairstyle and outfit stay as in @Image1 until the last frame. The background stays put.
How to phrase it
Same as with I2V: write what the result should be (“hands keep a natural shape”).
Why not second by second. The AI takes the sequence and timing of the motion from @Video1. If you repeat them in the prompt and the numbers are even slightly off from the reference, it ends up with two conflicting instructions, and the motion starts to stutter. The prompt doesn't retell the dance. It explains how to perform it.
Example breakdown
@Image1 repeats the dance from @Video1. Beautiful, realistic, high quality.
No phases, no hands, no expression, no camera. The AI decides on its own what matters in the reference and almost always loses the details: a gesture turns into a flail, a smile into a grimace, the camera drifts.
Timing and order come from @Video1. The prompt describes what the motion should be like: smooth hip sways, soft arms never above the shoulders, relaxed fingers. Separately: expression and gaze, range “slightly softer than the reference,” camera locked on a tripod, and her look and outfit staying as in @Image1.
The AI has both the sample and the instructions. It doesn't have to guess what matters: everything that matters is named.
@Image1 is the woman in the video. She performs the dance from @Video1, following its rhythm, timing and sequence of moves. Body: smooth hip sways on the beat, relaxed shoulders, soft fluid arm movements; her hands stay close to the body and never rise above shoulder level. Hands keep a natural shape with relaxed fingers throughout the movement. Head and face: light head tilts with the music, relaxed playful expression, soft smile, gaze mostly into the camera. Tempo and amplitude: same tempo as @Video1, movements slightly softer and smaller, smooth transitions between moves. Camera: locked off on a tripod at chest height, completely static for the whole video — no pan, no tilt, no zoom, no dolly, no sway, no horizon drift. Her face, body proportions, hairstyle and outfit stay exactly as in @Image1 until the last frame. The bedroom background stays still, bright neutral daylight from the window.
What a good reference looks like
A strong prompt won't save a bad sample. Pick your reference as carefully as you write your prompt.
One person in the frame
One continuous shot, no cuts or transitions
The person stays in frame the whole time, body and head never cut off
The camera stays still, if possible
Moderate pace: sharp and very fast moves transfer worse
Trimmed to just the part you need beforehand: extra seconds only get in the way
What a good start frame looks like
The frame and the reference should be built like they're “from the same scene.” The further apart they are, the more the AI has to make up, and the worse the proportions break.
The easiest guide is the reference itself: FBL APPS → FRAME PICK pulls sharp frames out of a clip. Build your start frame to match the framing and pose of the first one.
Start frame: a waist-up selfie. Reference: a full-body dance shot from three meters.
There are no legs in the photo, so the AI will draw them in itself. The head comes out too big, the height looks off, and the face warps on the very first turn.
Start frame: the model full-body, camera at chest height from three meters, standing in the same pose the dance starts with. There's room around her for arm movements.
Same framing, similar pose, similar proportions. All that's left for the AI is to transfer the motion, following your instructions.
I2V or Motion Control
It's not “better” vs. “worse.” They're two tools for different jobs. The choice depends on what exactly needs to happen in the clip.
What to check in the finished clip
The face in the last second matches the first
Clothes don't change their cut, number of buttons or layers
The background doesn't “breathe” or drift
The camera behaves logically: still or with a handheld shake, not jerking around
Body proportions don't change during movement
The clip looks fine on a phone at normal speed, the way a viewer will see it
Selection beats generation
Generating is half the job. The other half is honestly trashing whatever “almost worked.”
With people who actually make money in this niche, I see one underrated skill again and again: being able to throw away an “almost good” shot.
Out of ten results, usually one or two deserve the feed. The rest is material to learn from, not to post. And only the best photos move on to video.
Sorting is easier in batches than one file at a time in a folder: FBL APPS → PICKER gives you “keep / reject” on the keyboard and archives your picks in one click.
“Almost fine” is a no. The viewer doesn't give you credit for how many tries it took. They see one shot, and that shot either holds the image or breaks it.
Pre-post checklist
Is it her? Next to the reference face, it's the same person
The signature mark is there
It's clear who shot it and where the light comes from
Hands, jewelry, text, reflections: nothing weird
Doesn't look like an ad or a studio render
Fits the feed: same style, same vibe, same story
Passed the three-second test
Even one “no” and the shot doesn't go out. No “but it's so pretty” exceptions.
The AI executes. You make the calls.
Three formulas: face + prompt, photo + motion prompt, frame + reference + prompt. A good reference face going in, a prompt about the scene instead of ratings, motion step by step and honest selection at every stage. These aren't “secret prompts.” It's a way of thinking that separates a model people follow from a feed people scroll past and forget.
From formula to results
You now have the three formulas and the spots where each one breaks.
Someone who just mashes “generate” never sees any of that.
But knowing the formula isn't a skill yet.
Reading that a prompt should describe the angle and the camera is easy.
Writing a prompt that keeps the face unchanged to the last frame and the hands intact takes hundreds of generations and dozens of breakdowns.
The formula itself makes sense on the first read. The details are what eat your time: why the face changes at this exact angle, which word in the prompt breaks the hands, why the motion from the reference comes through in jerks. You can't see a mistake like that from the inside, so a single one can cost you weeks, and the AI looks like the one to blame.
The skill is built on this loop: make it, review it, fix one thing, and go again until the shot starts coming out without a fight. The loop spins fastest when your shots get reviewed by someone who's already run it.
You can start practicing everything in this guide today: one reference face, one shot by the formula, one short clip. Don't wait for the perfect moment or the perfect tool. They're not coming.
That's the whole foundation, free: how a shot is built, where it breaks and how to fix it.
Nobody can make that first shot by the formula for you. That's where the skill starts.
Being able to make a good shot is worth nothing if a month later you're burned out, you've lost your reference face and you can't remember what worked. How to turn all of this into a daily workflow is in the next guide.
Want to go further?
Stuck on a step? Drop your reference face, prompt and result in the Discord. Fresh eyes catch that kind of mistake faster than your tenth attempt.
Shot and prompt breakdowns, my personal experience and people doing the same thing you are: it's all there.
Copying, rewriting or repackaging without the author's consent is prohibited.