AI Filmmaking Program · Course 2 · Track B
Four lessons to turn the world’s most famous generator into a precision instrument: you calibrate, direct, edit, and repeat the result on demand.
Lessons
Leave knowing how to adjust the settings that separate a realistic photo from an illustration—and prove the difference in an experiment.
Generate a portrait that looks like an unretouched photo — asking for pores, asymmetry, and real human imperfections.
Expand the canvas, swap objects, and animate the frame—fixing the image instead of regenerating everything.
One scene, four adjustments: the final test that brings structure and parameters together in a method of your own.
Course 2B · Lesson 1
By the end of this lesson, you’ll adjust Midjourney’s controls intentionally — and prove what each setting does using the same prompt with three different configurations.
In professional image work, results don't come from luck — they come from adjustment. Most people use Midjourney straight out of the box and accept what it produces. Those who adjust four simple controls get to choose between photo and painting, between compliance and creativity. That's the control you gain today.
watch this lesson on video (English · optional)
↓ role to study
Midjourney comes out of the box with its own opinions: it beautifies, stylizes, and rounds things out. That helps with concept art; it gets in the way of realistic commercial work. The professional shift is understanding that each result is the sum of adjustments — and that they’re all within your reach.
Think in terms of four controls: the model (the engine version), the level of artistic intervention, how closely it follows the text, and the degree of variation among the four images in each generation. Adjusted together, they determine whether you get a catalog photo or a fantasy poster.
The social media manager working with a cosmetics store felt this on her first job: with everything set to factory defaults, the face cream came out in fairy-tale bottles. Two buttons adjusted later, it was a real package on a marble countertop—approved without revisions.
Before the technical names, here’s the on-ramp: these controls appear on the tool’s settings screen as options with English names—you only need to know which one to turn on and when; the menu does the rest.
Stylization (the control called Stylize): it’s the volume of the AI’s artistic opinion. Low (50–100), it follows instructions and the result leans toward a photo. High (600+), it becomes an artist—great for concepts, terrible for real products.
Raw mode (Raw): turns off aesthetic post-processing completely. It's the mode for anyone who wants strict adherence to the request — the right choice when the photographer needs the client's reference followed to the letter, without unwanted "improvements."
Common mistake
Increase stylization, hoping for “more quality.” More control doesn't increase quality — it increases the AI's opinions. For realism, the way to go is the opposite: low stylization, raw mode on, and quality comes from the description.
Midjourney comes with a monthly allowance of fast-generation hours — and, on higher-tier plans, an unlimited Relax mode that just waits in the queue. If you don’t understand the distinction, you’ll burn through your allowance on pointless tests and have none left when it’s time to deliver.
The professionals’ rule is simple: explore in relaxed mode, finish in fast mode. Drafts, variations, and experiments go into the queue at no cost; fast mode is reserved for the final polish, when every minute matters.
The owner of the neighborhood café who makes his own posters organizes the week this way: he lets ideas run in relaxed mode on Monday and saves the fast hours for Thursday, when he makes the weekend promotion poster.
The tool’s most underrated feature: in the personalization section, you rate a sequence of images (about 200 “like/dislike” choices), and Midjourney creates a style filter just for you—a personal style code you can use with your prompts.
The practical effect: results start leaning toward your aesthetic even before you describe it. For the photographer, that means a visual signature; for the social media manager, consistency across posts without repeating style paragraphs in every prompt.
The calibration experiment (you repeat it in practice)
Practice now 0/4 done
Leave with a series of 3 stylizations to compare and your settings written down—in ~12 minutes.
Adjusting the controls won’t break anything: every setting can be restored to its default with one click, and each generation creates a new file. Use relaxed mode, if available, so you don’t spend your fast hours on this test.
Ultra-photorealistic portrait of <seu assunto: uma pessoa, um produto, um prato>, natural skin texture with visible pores, soft window light, neutral colors, shot on full-frame camera --v 7 --style raw --stylize 100
You tuned the engine’s settings and know what each control does—exactly what this lesson promised.
Summary
Course 2B · Lesson 2
By the end of this lesson, you’ll build a portrait prompt layer by layer—skin, light, variation—and compare the out-of-the-box result with the real raw result.
A wedding client, a health insurance client, or anyone who needs a believable photo today can spot the “AI look” from a mile away: waxy skin, advertising-style lighting, overly perfect symmetry. That can cost you the job before the conversation even starts. The good news is that realism doesn’t depend on luck or a magic prompt—it depends on deliberately asking for imperfections, in layers, one at a time.
watch this lesson on video (English · optional)
↓ role to study
Ask for a “realistic portrait” with nothing else, and the engine returns smooth skin, advertising-style lighting, near-perfect symmetry — the factory default is beautiful, not real. Real people have pores, visible veins at the temple, one side of the face slightly different from the other. Asking for these details on purpose is what separates a stock photo from a photo that looks like it came from a real camera.
The restaurant owner tested asking only for “a photo of the chef preparing the dish”—and got a margarine-ad chef with wax-doll skin. The turnaround came when he started naming the layers: visible skin pores, slight facial asymmetry, real kitchen lighting.
Before
"A photo of the chef preparing the dish." No layers — AI decides on its own how much realism to deliver (not much).
After
"A photo of the chef preparing the dish, with a focus on biological realism, visible skin pores, subtle veins at the temple, natural facial asymmetry, soft ambient light, unpolished skin texture." Each layer named, one by one.
The payoff: the same dish, the same chef—the second version passes the “does this look like a real photo?” test.
Before showing you how to write it, here’s the ramp-up: think of a request to a seamstress—you first name the garment (a dress), then the finish (hem, type of fabric), and only at the end the style (classic or modern). A realistic portrait request follows the same order: first the subject, then the biological details, and finally the style and photographic technique.
A loose poetic paragraph ("a beautiful person in the golden light of dusk, with a gaze that tells stories...") is lovely to read, but the engine doesn’t know which part deserves the most attention. A layered structure solves this by assigning a fixed place to each type of information.
The photographer reused the studio checklist she already had before AI: subject (who, what), details (skin, texture, light), style (lens, development). It became the outline for every portrait prompt she wrote, without reinventing it for each new client.
The settings screen has two extra controls that promise “more variety”: Weirdness and Variety. Think of them as recipe seasonings—a little too much and the dish doesn’t become more interesting; it becomes strange.
Weirdness pushes the scene toward the surreal: a hand with an extra finger, an eye out of place. For a realistic portrait or product shot, keep it at zero. Variation controls how much the four options from each generation explore different framing — 15 to 25 already gives you variety without straying from the requested subject.
A social media manager tried high Weirdness on a skincare routine photo, thinking it would “bring the scene to life”—the result was a hand with an extra finger, almost invisible in the small preview, glaring in the enlarged post. Setting Weirdness to zero fixed it that same day.
Common mistake
Increase Weirdness, thinking it gives you “more real human variety.” This control adds surreal deviation, not natural variation — hands, faces, and proportions start to break down. For a realistic portrait or product, keep it at zero and use Variety when you want to explore different framings.
Once the image looks good, there’s still one step left that sounds technical but is simple: increasing the size and detail of the final file. In the tool, this step is called upscale — enlargement.
There are two expansions with opposite purposes. The smooth preserves the structure and skin texture exactly as they came out — it's the right choice for photos that need to keep looking real. A creative adds artistic detail on its own — great for posters and concept art, risky for products or realistic portraits, because it may “improve” the scene until it’s no longer the same.
The restaurant owner uses subtle upscaling on the dish photo for the menu—it needs to keep looking like that dish—and creative upscaling on the poster for a themed event, where extra atmosphere is welcome.
The final professional step isn’t accepting the first good image—it’s refining one specific detail without regenerating everything, using the region variation tool, and only then choosing the right upscale for the image’s intended use.
A social media manager uses this workflow every week: she generates the influencer’s portrait with the layered formula, adjusts only the light on one side of the face with the region variation tool, and upscales in gentle mode before posting—without ever regenerating the whole scene over one detail.
The finishing workflow (you repeat it in practice)
Practice now 0/4 done
Leave with two versions of the same portrait—with and without the realism layers—and the right enhancement selected, in ~14 minutes.
Nothing here is final: each try creates a new file, and the original image is never overwritten. If the regional variation scrambles a detail, repeat the prompt and state more clearly what should stay the same.
Macro photography portrait of <seu assunto: um chef, um artesão, uma cliente>, focus on biological realism, visible skin pores, subtle veins on temples, natural facial asymmetry, non-glossy skin texture, soft ambient light, shot on 85mm lens, f/2.8, unretouched --v 7 --stylize 200 --weird 0
You assembled a layered portrait, compared it with the factory default, and chose the right upscale—exactly what this lesson promised.
Summary
Course 2B · Lesson 3
By the end of this lesson, you’ll fix an image without recreating the whole scene—expand the frame, replace an object, and lock in a character or style for new scenes using the tool’s editor.
The first generation rarely gets everything right—usually just one thing. Starting over until that one thing is right wastes time and luck at once. There’s a shorter way, one image professionals used before AI: fix only the part that’s wrong and leave the rest untouched. That’s the lab you’re opening today.
watch this lesson on video (English · optional)
↓ role to study
An almost-perfect image is the easiest to ruin—the instinct for beginners is to ask "generate it again," and the whole good scene disappears, including the parts that were already right. The tool’s editor does the opposite: you point to only the problem area (the light, an object, the frame boundaries), and everything else stays exactly as it is.
The photographer lost one of the week’s best portraits this way: she asked, “generate it again, but without that pole in the background”—and got an entirely new scene, with a different pose, lighting, and client expression. The pole was gone, but so was the good portrait.
Common mistake
Asking for an entirely new generation to solve a specific problem. A new generation doesn’t “edit” the previous one — it recreates the scene from scratch, including the parts that were already right. Use the editor to mark only the area or feature that needs to change; the rest of the image stays locked.
The editor screen looks like a professional editing program packed with buttons—but the canvas expansion feature works like stretching a tablecloth: you pull the edge outward, and the missing fabric continues the same cloth; it doesn’t turn into a visible patch.
The generative fill tool lets you drag the image boundaries outward — from a square to a wide cinematic frame, for example — and AI rebuilds the surrounding environment while preserving the original scene’s lighting, perspective, and style.
The social media manager received a square photo that was perfect for the feed, but also needed a wide cover image for the client’s video channel. She expanded the frame to the sides: same scene, same light, just more breathing room around it. No reshoot.
There’s a feature that works like an anchor: the Universal Reference (Omni Reference, on screen). You upload an image that's already finished and tell the tool "keep this" — it could be a person's face or the entire visual style of a photo. An intensity control determines whether the likeness is exact or just inspired by it.
The effect is the same in two very different situations—and that’s where the feature shows its strength: the same anchor, two uses.
The same constraint, two uses
The selection tool works like a real brush: you paint over the object you want to replace (a chair, a product, a background) and write only what should appear in its place — not the whole scene again.
The social media manager received stock photos with a competitor’s product on the table. She selected just the product with the brush and wrote her client’s packaging in its place — the rest of the scene (table, light, hands) stayed intact, with no new photo shoot.
The editor’s final module turns a still image into a short video. Think of it as the wind control for a storefront display: set it low, and only the fabric moves; set it high, and the whole scene comes to life.
You choose between low intensity (subtle movement—hair, breathing, smoke) and high intensity (wide, more dramatic movement)—and you can even define the first and last frames to control exactly how the scene transforms.
The social media manager animates the hero image for a product relaunch for the client’s short video, at low intensity—just the packaging subtly gleaming in the light—a 4-second clip that looks filmed, made from a single photo.
From a still frame to a clip (you’ll practice it yourself)
Practice now 0/4 done
Walk away with one of your images expanded or with an object swapped, and a documented decision about which tool solved the problem—in ~12 minutes.
The editor always works on a copy—the original image is never deleted, and each attempt creates a new file. If the result looks strange, that’s test data, not your mistake: undo it and try again.
You fixed one of your own images without regenerating the entire scene—exactly this lesson’s promise, fulfilled with your own photo.
Summary
Course 2B · Lesson 4
By the end of this lesson, you’ll generate the same scene four times, changing one parameter at a time—and leave with your own documented method for which setting matters most for each type of work.
So far, you’ve learned individual pieces: stylization, raw mode, realism layers, and the editor. The most practical question remains — when someone asks for “more dramatic” or “more on-brand,” which setting do you adjust first? This lesson wraps up the course by testing the parameters side by side in the same scene until the answer becomes instinctive.
watch this lesson on video (English · optional)
↓ role to study
Back at the beginning, with a director’s eye: every strong prompt names five elements, in order — the subject, the action (what they do), the camera (the shot and lens), the setting, and the scene’s emotional tone. It’s the same discipline as a commercial script, just written in one line.
Social media stopped writing "photo of a beautiful cake" and started writing piece by piece: subject (the three-tier cake), action (pastry chef adding the final detail), camera (medium shot, 50mm lens), setting (professional kitchen, morning light), mood (focus, homey warmth). The five-sentence request became the standard for every client campaign.
The settings screen has more controls than Stylize and Raw Mode: Chaos (the distance between the options in each generation), Quality (the render’s level of finish), and Style Reference (which borrows the visual identity of an existing image). Think of them as separate measuring tools—each measures just one thing.
Chaos controls how far apart the four options in each generation are from one another — low for stable commercial realism, high for exploring very different ideas. Quality controls the render’s level of finish — low works for a quick draft, high for the final delivery.
The photographer wanted to know the real effect of high Chaos—but changed Chaos and rewrote part of the prompt at the same time. The scene came out so different that she couldn’t tell whether the parameter or the text had changed everything.
Common mistake
Change the prompt text and one parameter at the same time, in the same test. Without isolating the variable, you don’t know which of the two caused the change—the test becomes a guess. Lock the prompt text, change only one parameter at a time, and compare the generations side by side.
There’s a parameter that works like a brand label: the Style Reference (sref on screen, short for style reference). You point to an image that already has the right look, and every new prompt follows that same visual identity — lighting, color, finish.
The restaurant owner had a photo of the opening dish that everyone complimented. He used it as a style reference to photograph the other twelve dishes on the new menu—all with the same warm light and finish, without writing the style description twelve times.
Test yourself
You’ve wrapped up a campaign and want the next 8 images to look like they came from the same photographer, with the same lighting—but each featuring a different product. Which setting does this without making you rewrite the style description for every image?
The final proof that you’ve mastered the controls isn’t memorizing numbers—it’s running the same scene four times, changing only the parameter combination, and naming what each version achieves: a precision test (commercial realism), a cinematic test (controlled drama), an exploration test (free variation), and a brand identity test (with a style reference).
This is the workflow the social media manager runs before finalizing any new content package: the same base scene, four adjustments, and they choose which combination will become that client's standard.
The four-settings test (you’ll repeat it in practice)
Practice now 0/5 done
Leave with up to four versions of the same scene and a written conclusion about which setting had the greatest impact on your type of work—in ~15 minutes.
Each generation is a new, independent file — no previous version is lost, and you can compare them all side by side at the end without redoing anything.
<seu assunto>, <ação em uma frase>, medium shot, 50mm lens, <ambiente: onde e que hora>, <clima emocional da cena> --v 7 --chaos 10 --stylize 100 --q 2
You ran the same scene with different settings and know which one to use for each type of request—the method that wraps up this entire track.
Summary