AI Filmmaking Program · Course 2 · Track A
Seven lessons to master the frame: you’ll start directing every image—shot, lens, light, and character—instead of accepting whatever the AI gives you.
Lessons
Write prompts the way a director describes a shot — and watch the same scene change level instantly.
Angle sheet, relighting, production sheet, and master layer—fine control without guesswork.
Understand why AI changes faces—and create the reference sheet that reduces variation.
Define your actor once, train the angles, and continue the same scene without losing their identity.
Place, action, emotion: the visual grammar that needs no words.
The eight steps that take a raw idea to a polished hero frame — the whole system in one of your scenes.
Turn everything that worked into reusable templates—and never start from scratch again.
Course 2A · Lesson 1
By the end of this lesson, you’ll write image prompts using the structure directors use—and see the difference in a side-by-side comparison.
The difference between just any image and a film frame rarely comes down to the tool. It comes down to the order and precision of the prompt. Studios already use image generators to develop films and campaigns — and they all write the same way: as if describing a film shot.
watch this lesson on video (English · optional)
↓ role to study
Tools like Nano Banana—a fast image generator that’s good for visual development—have been trained on decades of film and photography. They recognize industry terms: wide shot, close-up, 50mm lens, backlight. Those who speak this language get scenes; those who don’t get generic illustrations.
And here’s the training principle: you don’t need this specific tool. The structure in this lesson works with any generator — the language of cinema is the same in all of them.
The social media manager feels the difference on the first test: "perfume photo" returns a shallow catalog shot; "close-up of the bottle, macro lens, golden side lighting, dark background" returns a magazine ad.
A shallow prompt describes the subject. A director’s prompt describes the shot.
Film directors and cinematographers always describe a shot in the same order—and your prompt will follow it:
The photographer recognizes her own mental set checklist. The new part is that the checklist now becomes text—and the order matters because the generator gives more weight to what comes first.
The fastest way to master the structure isn’t memorization: it’s a controlled experiment. You generate the scene, change a single piece — just the lens, or just the lighting—and generate again. The side-by-side comparison shows what that element controls, without any theory.
The restaurant owner tested this with Saturday’s feijoada spread: he changed only “midday light” to “late-afternoon light coming through the window” and saw the same dish go from cafeteria food to a magazine restaurant. One element, a huge leap.
Before
"A photo of a pastry chef decorating a cake." Subject described, shot left up to AI.
After
"Medium shot, 50mm lens, morning kitchen with window light, pastry chef focused on applying the icing, soft backlight, realistic film style." The six pieces, in order.
The payoff: the same tool, the same scene—and the second image looks like a film still, not a stock photo.
Test yourself
You want to find out in practice what a lens changes in a scene. What’s the right experiment?
Practice now 0/3 done
Leave with two versions of the same scene, changing just one element, and a conclusion about which element had the greatest impact—in ~10 minutes.
Text and generation only: none of your files are touched, and no attempt is lost. A bad image is experiment data, not a failure.
<tipo de plano: wide shot / medium shot / close-up>, <lente: 35mm / 50mm / 85mm lens>, <ambiente: onde, hora do dia, clima>, <ação: o que o personagem faz, em uma frase>, <luz: direção e humor — golden hour, luz de janela, contraluz>, <estilo: cinematic film still, realistic textures, natural colors>
You directed the same scene twice and can explain what each element controls—that’s exactly what this lesson’s promise called for.
Summary
Course 2A · Lesson 2
By the end of this lesson, you’ll fix the lighting in an image without recreating the scene—and apply the quality layer that improves every prompt you write from now on.
Generating a good image once is luck; generating it again on demand is control. The difference comes down to a small set of techniques professionals use every day—four levers you pull as needed instead of leaving things to chance.
watch this lesson on video (English · optional)
↓ role to study
A beginner types and waits; a professional decides and checks. Between the two is a workshop principle: every adjustment changes one thing at a time, leaving everything else intact. That’s how you fix a scene without destroying it.
The four levers in this lesson follow that principle: the angle sheet only the point of view changes; the relighting only the light changes; the production profile organizes the complex request; the master layer elevates the finish of any generation.
The restaurant owner understands this instinctively: when the broth is good but needs salt, no one throws out the pot—they adjust the salt. In the next screens, you’ll learn how to “adjust the salt” in your images.
You already know the idea from course 1; here, it becomes a study tool. The contact sheet shows the same scene frozen in time view from several angles: same character, setting, and lighting — only the camera changes position.
Professional use goes beyond consistency: this is how you studies one scene before deciding on the final shot. Instead of imagining what it would look like from above, below, or the side — you see all nine options on one sheet and choose with your eyes.
The photographer uses the sheet the way she used to use a shoot sketch: to figure out that the buffet table looks better photographed from low down, at glass height, than from above—before spending the shot that counts.
The scene came out perfect, but too dark. A beginner’s instinct is to generate everything again — and lose the good scene. The right lever is the relighting: you upload the finished image and ask for the change only in the light.
The prompt names what changes and locks everything else: "keep the composition, people, and objects exactly as they are; only brighten the scene with soft late-afternoon light coming through the window on the left."
The restaurant owner saved the best photo of the dining room this way: the right setting, customers well positioned, but basement lighting. After relighting it, the same scene looked as if it were lit by huge windows—without moving a chair.
Common mistake
Asking to “improve this image.” A vague prompt gives the AI permission to redo everything — and the good scene dies along with the bad lighting. Name what changes (only the lighting) and explicitly lock what stays.
A complex scene — multiple elements, precise camera, specific mood — overwhelms a rushed prompt: something always gets lost in the paragraph. The professional solution is the same as on a real set: one production profile, with a field for each decision.
Before you see the example: this is a written form—each line has a label (camera, character, setting, light) and your answer. The text format is called JSON, and the strange punctuation (braces, quotation marks) is just how the machine separates the fields. You fill in the answers; copy the rest as is.
{
"cena": "chef solitário na cozinha antes do serviço",
"camera": { "plano": "wide shot", "angulo": "low angle", "lente": "35mm" },
"personagem": { "aparencia": "chef de meia-idade, avental de linho",
"acao": "acende a primeira boca do fogão", "emocao": "calmo e alerta" },
"ambiente": { "local": "cozinha profissional vazia", "hora": "antes do amanhecer",
"detalhes": "vapor sutil, panelas de cobre penduradas" },
"luz": { "estilo": "uma única luz quente sobre o fogão", "paleta": "tons âmbar e aço" },
"acabamento": "ultra photorealistic cinematic, 8K"
}
The social media manager discovered the hidden benefit: the profile is ready to present. The client approves it field by field before of the generation — and changes become one-line edits, not meetings.
The final lever is the simplest: a fixed list of finishing terms—natural film lighting, realistic reflections, subtle grain, visible pores, precise focus—that you paste into the end for any request. It doesn't change the scene; it changes the level of polish.
ULTRA PHOTOREALISTIC CINEMATIC SCENE, NATURAL FILM LIGHTING, GLOBAL ILLUMINATION, REALISTIC REFLECTIONS, KODAK CINEMATIC COLOR GRADING, SUBTLE FILM GRAIN, HIGH DYNAMIC RANGE, SHARP FOCUS, CINEMATIC DEPTH OF FIELD, REALISTIC TEXTURES, NATURAL SKIN PORES.
The photographer thinks of it as the finishing touch she applied to every good print in the darkroom: a house quality standard, not a new decision for every photo.
Using the master layer
Practice now 0/3 done
Walk away with one of your images relit without losing the scene, and a comparison with and without the master layer—in ~12 minutes.
The original image stays saved; relighting always creates a new file. If the AI changes something that should stay still, repeat the request and emphasize what’s locked.
Keep the composition, characters and objects exactly as they are. Change ONLY the lighting: <descreva a luz nova — ex.: soft warm late-afternoon light coming from a window on the left, gentle shadows, natural exposure>. Do not add or remove anything.
You fixed the lighting without sacrificing the scene and proved the effect of the quality layer—the two most commonly used everyday controls, mastered.
Summary
Course 2A · Lesson 3
By the end of this lesson, you’ll create your character reference sheet—the document that reduces face variation in every scene you generate from now on.
Every professional production visualizes the scene before spending a cent filming it. With AI, this dress rehearsal takes minutes—but it runs into a limit nobody tells you about: AI forgets the face from one image to the next. This lesson explains why and gives you the classic defense professionals use.
watch this lesson on video (English · optional)
↓ role to study
In film, storyboards exist to save money: getting it wrong on paper costs nothing; getting it wrong on set costs the whole crew a day. A storyboard determines the composition, character position, angles, and atmosphere before any camera starts rolling.
With a fast generator like Nano Banana, this step—which studios call previsualization—is within anyone’s reach: you can test compositions, lighting, and framing in minutes, like sketching ideas on a napkin that talks back.
The photographer uses this in her sales pitch: instead of describing the shoot over the phone, she shows three preview frames of the concept—and closes the deal on the spot, because the client saw it before paying.
Mistakes on paper are free. Mistakes in production cost you the day.
Here’s the honest limit of the technology: the generator creates each image independently, with no memory of the previous one. That’s why the face changes a little, the hair shifts, and the age fluctuates—even with the same description.
Write down the practical takeaway, because it guides everything in this learning path: professional work with AI aims to reduce variation, don't eliminate it. Anyone who promises perfect consistency is selling; anyone who reduces variation until no one notices is doing the work.
The restaurant owner experienced this in the series of posts about the fictional chef: between the second and fifth images, the chef aged ten years. It wasn’t a flaw in the tool—it was the missing safeguard introduced in the next step.
The professionals’ solution predates AI—it comes from animation studios: a character reference sheet, with the same person in four official angles: front, three-quarter, profile, and back.
This sheet becomes the character’s identity document: you send it along with future prompts, and the generator has a visual reference to consult instead of reinventing the character from memory.
The social media manager keeps the sheet in the client’s folder, next to the brand guide—because that’s what it is: the character’s identity guide.
Before
Without a sheet: every scene recreates the face from memory—five posts, three different people.
After
With the reference sheet attached to the prompts, the generator checks the guide—the variation drops to the point where the audience won’t notice.
The payoff: a document generated once protects every future scene with the character.
With the sheet in hand, the professional workflow has four stages: fixed description of the character (written once, never changed) → reference sheet generated → the same description repeated in everyone the requests → and the shots generated in cinematic order: wide to establish the setting, medium for the action, close-up for the emotion, reaction shots to tie it together.
Notice the principle behind it: disciplined repetition. Anything that can be fixed—the description, sheet, and shot order—gets fixed; creativity lives in what the scene tells, not in reinventing the character.
The photographer compares it to a maternity portfolio: the pose sequence has been the same for years—wide shot, medium shot, detail of the hands. Following the sequence is what frees her attention for the client’s emotions.
Test yourself
Why does the character’s face change between two generations with the same description?
Practice now 0/3 done
Leave with a reference sheet of your character from four angles—in ~12 minutes.
Nothing here changes your previous images; the sheet is a new document. If an angle looks strange, generate it again—the official character is the one you approve.
Create a character reference sheet with the same character shown from four angles on a clean neutral background: front view, three-quarter view, side profile, and back view. The character: <cole aqui a descrição fixa do seu personagem — a mesma da aula 5 do curso 1, sem mudar uma palavra>. Keep identical face, hair, outfit, colors and proportions in all four views. Neutral even lighting, no scenery, no props, no text.
Your character now has an ID: a fixed description and a four-angle sheet—consistent facial features are no longer a matter of luck.
Summary
Course 2A · Lesson 4
By the end of this lesson, you’ll generate a sequence of shots from the same scene, with the same character, using the complete locking system—define, train, continue, expand.
The reference sheet from the previous lesson reduces variation — but by itself, it can’t hold an entire sequence together. What’s missing is the system around it: a description that never changes, angle training, and the discipline to continue the scene instead of starting over. That’s what separates unrelated images from a scene that holds together.
↓ role to study
The system starts with a producer’s decision: the character is defined just once, in writing, and that text becomes law. Face, hair, clothes, accessories, tone—everything recorded in the fixed description. Change a comma halfway through the project, and the system breaks.
Consistency comes from combining four supports: a fixed description, reference images, the same repeated prompt structure, and an identity control phrase in every prompt ("same face, same clothes, don’t redesign").
The social media manager learned this from the fictional veterinarian series for the pet shop: in the third post, she “improved” the description—and the veterinarian became a different woman. Since then, the description has lived in a locked document, and she only copies and pastes it.
Before any scene, the model needs to know your actor from every angle. You generate the training set: front, profile, back, and two close-ups—all with the fixed description and a neutral background.
It's the lesson 3 worksheet taken to a professional level: the close-ups are included because a real scene calls for a large face — and a large face without practice is where identity slips the most.
The photographer thinks of it as the camera test you do with a new model before the paid shoot: half an hour of controlled angles that saves the whole day.
Common mistake
Skipping the practice and going straight to the scene. It works until the first different pose—then the face breaks and you’re back to square one. Ten minutes of angle practice costs less than redoing an entire sequence.
Here’s the mindset shift at the heart of this lesson: stop generating new images and start continue the same scene. The request changes from "create an image of..." to "continue this scene: now the camera moves closer to the face; nothing else changes".
Small changes between shots are the secret to continuity: AI has little room to make things up, and the viewer experiences one scene—not a collection of shots.
The restaurant owner put together the wood-fired oven sequence this way: a wide shot of the dining room, continuing with the camera moving toward the counter, continuing with a close-up of the pizza coming out. Three prompts, one scene—and the same oven in every shot.
With the base scene established, comes the payoff: expansion. Extend the environment beyond the frame, increase the scale, add weather effects—rain, dust, fog. The rule that governs it all: don't change the character; change the camera, the light, and the environment.
It's the same soap-opera logic social media managers know from the other side of the screen: the actor stays the same all year; what changes are the sets, the episode's costumes, and the cinematography.
The lock system, in five steps
Practice now 0/4 done
Walk away with a base scene + a continuation + an extension, all with the same character—in ~15 minutes.
Each prompt generates a new file; the base scene stays saved even if the continuation comes out wrong. Did the identity drift? Strengthen the control phrase and repeat only that shot.
Continue this exact scene. Camera change ONLY: <a mudança — ex.: move closer to a medium shot of the character>. Keep the same character, face, hair, outfit, environment, lighting and color palette. Same identity as the reference images. No redesign, no new elements. Stable face.
You directed an entire scene with the actor locked—the system that turns loose images into a film sequence.
Summary
Course 2A · Lesson 5
By the end of this lesson, you’ll create a sequence with just three prompts that establishes the setting, shows the action, and reveals the emotion—without writing a single caption.
An image on its own can be beautiful and still say nothing. Most AI prompts produce isolated frames with no connection to one another — and what separates a standalone image from a movie scene is precisely this sequence grammar, which anyone can learn in one lesson.
↓ role to study
The shortest sequence that tells a complete story has three links, always in this order. The wide shot sets the scene: shows where the scene takes place, the scale of the location, the mood of the moment. The medium shot shows the action: what the character is doing, the movement that drives the story forward.
The third link is the close, which reveals the emotion: what the person feels, without needing a caption. Together, the three in the right order replace any explanatory text.
The photographer uses this sequence at the quinceañera: a wide shot of the decorated hall sets the scene; a medium shot of the birthday girl dancing with her father shows the moment; a close-up of her teary eyes closes the scene—three photos, no captions, the whole story told.
Place sets the scene. Action moves it. Emotion stays.
Camera height isn’t just framing — it’s emotional direction. A low camera looking up makes the character seem larger, dominant, in control. A high camera looking down makes that same character seem small, exposed, vulnerable.
It's the same technique cinema has used for decades to show, without saying a word, who's in control in a scene — and you request this angle directly, in the same prompt you learned in lesson 1.
The restaurant owner used a low angle for the new menu cover photo: seen from below while presenting the signature dish, the chef looks confident and in command of the kitchen—exactly the effect the launch called for.
Test yourself
You want the social media manager to look small and overwhelmed in a post about Monday’s behind-the-scenes chaos. What angle should you ask for?
A scene looks flat when everything is at the same level of sharpness — like an ID photo. A scene looks like a film still when it has layers: something blurred very close to the camera, the subject in focus in the middle, and the background softened behind them.
Asking for depth is simple: describe what’s in the foreground (even if blurry), the main subject in the middle, and what’s in the background. The generator arranges the sharpness of each layer on its own.
The social media manager applied this behind the scenes at the beauty salon: blurred scissors in the foreground, the hairdresser in sharp focus cutting hair in the middle, and other clients softened as they waited in the background—the photo gained the depth of a soap opera still, not a candid phone shot.
Before
A photo of the salon with no depth: everything on the same focal plane, with the client and background competing for equal attention.
After
Same scene with three described layers: a blurred foreground, a sharp hairstylist, a soft background—the eye knows exactly where to go.
The payoff: the same scene, without changing the subject, gains the sense of depth you only get in a film still.
The last ingredient in a cinematic scene isn’t the subject—it’s the air between the camera and the subject. Light hitting dust, steam, or fog creates visible beams, and visible beams create atmosphere instantly.
The prompt is one extra sentence at the end: describe light coming in from an angle (a window, door, or crack) and what’s suspended in the air—fine dust, steam from a cup, or low morning mist.
The photographer used this in the newborn shoot: late-afternoon light streaming through the barn, dust suspended in the beam, the baby in the mother’s arms inside that golden shaft—the atmosphere did half the emotional work of the photo on its own.
Practice now 0/3 done
Leave with three prompts for your scene—wide, medium, close-up—ready to generate, in ~10 minutes.
Text only: nothing of yours is changed, and generating the three images afterward is optional for this exercise. If the close-up doesn’t convey the emotion, adjust just that phrase and try again.
PLANO GERAL — situe o lugar: <wide shot>, <ambiente e hora do dia>, <clima geral da cena> PLANO MÉDIO — mostre a ação: <medium shot>, <o que o personagem faz>, <ângulo: low angle (força) ou high angle (vulnerável)> CLOSE — revele a emoção: <close-up>, <expressão ou detalhe do rosto/mãos>, <luz e atmosfera: poeira, vapor ou neblina no ar>
You wrote the shortest sequence that tells a complete story—setting, action, and emotion, ready to become three images.
Summary
Course 2A · Lesson 6
By the end of this lesson, you’ll take one of your scenes through all eight steps of the complete workflow—from a blank world to a polished hero frame—without skipping any.
So far, you’ve learned individual pieces: how to request a shot, fix the lighting, lock in a character, and sequence three shots. On their own, they still depend on you remembering the right order for each new scene. This lesson gives you the system that ties everything together — the same discipline that keeps you from starting over from scratch on every project.
↓ role to study
So far, you’ve learned individual pieces: how to request a shot, fix the lighting, keep a character consistent, and sequence three shots. What’s missing is the piece that ties everything together — a fixed-order workflow, from scratch to a frame that’s ready to publish.
Film professionals don’t make up the order for every project: first they decide on the world, then the story, then they produce the image, and finishing comes last. Skipping this order is the most common cause of rework.
The restaurant owner felt the difference when he stopped generating standalone images for the feed and started running every campaign through the same workflow: fewer attempts, results closer to what he had in mind before he started.
This may look like a long list, but it works like a recipe: follow the order once, and it repeats itself for every new scene. The first four steps build the scene; the last four refine what’s already there.
The eight steps in the workflow
The social media manager runs this entire workflow in one afternoon for a client’s product launch: instead of generating ten separate images in the hope that one will work, she comes away with a sequence that’s ready for the carousel from the start.
The eight steps don’t require eight different tools. Most image generators today organize this in your image generator’s workflow environment: each step becomes a step on the same timeline, without switching programs.
Think of it as a production line: you put the raw idea in one end, and each step moves the scene forward without losing what was decided in the earlier steps.
The photographer uses this setup for the bridal campaign: the character (the bride) and the world (the ceremony garden) stay fixed in one tab, and she moves through the eight steps without rebuilding anything from scratch.
The most tempting shortcut is also the most expensive: applying atmosphere and polish before locking in the composition. The scene looks beautiful, but it's wrong — change one detail in the composition and all the finishing work is lost because it was applied to the wrong frame.
Common mistake
Polish before approving the composition. Finishing (steps 6 to 8) only begins once the hero frame (step 4) is locked. Approve the composition first; the rest can be redone in minutes, but the polish can’t.
The restaurant owner learned this the hard way: he added steam and film grain to the photo of his signature dish before noticing that the knife angle was wrong. Fixing the angle meant redoing all the polish—from then on, he locked the composition before touching the finishing details.
Here’s the hidden efficiency of the workflow: the first two steps — world and character — are done only once per project. Every new scene in the same project goes straight to step 2: story planning.
That means the fifth scene in a campaign is always faster than the first—the world already exists; only the story changes.
A social media manager sees this in practice: she built the fictional restaurant’s world once, in her first week on the job; in the scenes that follow, each new post goes straight into the plan—half the workflow, reused every time.
Practice now 0/4 done
Walk away with a scene that has gone through all eight stages, from worldbuilding to final polish—in ~15 minutes.
Each step generates a new file; previous versions remain saved. Stuck on a step? Go back to the previous one and adjust — repeating a step doesn’t lose anything.
You ran the whole workflow, from scratch to final polish—the same discipline that separates random attempts from real production.
Summary
Course 2A · Lesson 7
By the end of this lesson, you’ll have a personal library with at least three tested prompt templates — and proven you can reuse one in a new scene.
In the last six lessons, you wrote prompts that worked: the three-shot sequence, the quality layer, the identity control phrase, the production sheet. Without a place to keep them, each of those wins gets lost in the conversation—and the next scene starts from scratch.
↓ role to study
A case to get a feel for the problem: Renata, a freelance social media manager, spent fifteen minutes rewriting from memory a prompt for a client’s product carousel — the same prompt she had nailed, tested, and approved three weeks earlier, but never saved anywhere.
Over the last six lessons, you’ve written prompts that worked: the three-shot sequence, the quality layer, the identity-control phrase, the production sheet. Each of these is a model — text ready to reuse—but only if you save it.
Without a library, every good prompt gets lost in the conversation, and you reinvent the wheel with every new scene. With a library, you copy, paste, and adjust only the details.
This doesn’t require any new tools—the same notepad or text document you already use works perfectly.
Building your library
The restaurant owner organized his into four sections: the production sheet for the signature dish, the fictional chef’s control phrase, the quality layer, and the three-shot sequence—four templates, ready for any new dish on the menu.
A model is only worth anything if it works in a scene it’s never seen before. Testing a reuse means taking a saved model and applying it, without rewriting it, to a completely new scene.
If the result looks similar to what worked the first time, the model is validated—and can be used for any future scene of the same type.
The photographer tested it this way: she took the quality layer she used for a wedding shoot and pasted it, without changing a word, into a request for a completely different newborn shoot. The finish improved just as much in the new scene.
This library depends on no one but you. It grows exactly in step with your own work, one model at a time, without depending on anything else.
Each new project you make is a chance to test a saved template or create a new one. After a few weeks, most of your scenes will start with a copy from the library, not a blank page.
It's the same logic as a restaurant owner who keeps family recipes in a personal notebook: they don't wait for someone else to write the next one — they write it themselves the first time it turns out right.
Practice now 0/3 done
Leave with at least three saved templates and a reuse test completed on a new scene—in ~10 minutes.
No image or previous prompt is deleted—you’re only copying what already worked to a new place. Made a mistake copying it? The original is still where it was.
Your library has the first models saved and tested—starting now, no new scene begins from scratch.
Summary