Module 7 • Lesson 1🎛️ Generative at Pro Standard
Amateur AI is a lucky prompt; professional AI is control. You direct the model with reference images, structure, and a matched style so it serves your art direction instead of surprising you — and you generate into an existing scene, not from a blank void.
📚 What You'll Learn
By the end of this lesson, you will be able to:
- Direct generation with a reference image (composition & style)
- Constrain structure so the result matches a layout you decide
- Match an existing style and light so output fits a project
- Generate into a selection so it integrates with real pixels (Module 2)
- Treat generation as raw material you curate and finish by hand
⏱️ Estimated Time: 55 minutes
🎯 Project: A controlled generative element — driven by reference and structure — integrated and hand-finished into a real composite.
ℹ️ Note on this course's images
Generative results vary by tool, model, and version, so this lesson teaches the control workflow with diagrams rather than one-off generated examples. The principles — reference, structure, style, integrate, finish — apply to Firefly and any other generative model you use.
In This Lesson
👤 The Goal: Direct, Don't Gamble
Two ways to use a generative model. On the right, a bare prompt — you type words and take whatever the dice give, re-rolling until something is passable. On the left, you supply a reference and a structure, so the output lands on your composition, in your style, first time. Same model; the difference is control. Drag the handle:
🧠 Mental Model: Constrain the Model
A generative model will happily invent anything that fits your prompt — which is the problem. Left unconstrained it wanders in composition, style, and detail, so you re-roll endlessly hunting for a frame that matches what you already pictured. The professional inverts this: you remove degrees of freedom — pin the composition with a structure, pin the look with a reference, pin the palette to your project — until the model can only produce something close to your intent.
Think of it as directing a very fast, very literal concept artist. Vague direction gets you a hundred wrong pictures fast; precise direction — "this layout, this lighting, this palette, fit this hole" — gets you a usable one. And whatever comes back is a starting point you integrate (Module 2) and finish by hand, never the final pixel.
🎚️ Reference, Structure & Style
Modern generative tools accept far more than a text prompt. A reference image steers overall look and composition. Structure controls (a rough sketch, a depth or edge map) lock where things go, so the model fills your layout rather than inventing its own. A style reference matches palette, rendering, and mood to your project. Used together, the prompt becomes the smallest part of the instruction — the constraints do the steering.
⚠️ The model is confidently wrong about specifics
Generative output drifts on exact things — counts, text, logos, hands, product details, continuity. Constraints reduce it but never eliminate it, so audit at 100% and fix specifics by hand. Trusting the model on a client's product name or logo is how an embarrassing deliverable ships.
🧩 Generate Into the Scene
The most controlled generation happens inside an existing image, not on a blank canvas. Make a selection in your real composite and generate only there — the model samples the surrounding pixels, so the new content inherits the scene's perspective, light, and palette and lands far closer to integrated. This is generative fill (Intermediate Module 6.1) used as a compositing tool: grow connective foliage, extend a wall, add a distant element that already matches.
✅ AI accelerates; craft finishes
Everything you learned in Module 2 still applies to a generated element: match the edge, the grain, the depth of field, and lay one unifying grade. The model gets you a plausible starting layer in seconds; your integration and hand-finishing are what make it survive scrutiny. AI changes the speed, not the standard.
🛠️ Guided Build: A Controlled Generate
Open a real composite that needs one new element (a background structure, foliage, an object).
Step 1: Gather constraints · 8 min
- Pick a reference image for look, and rough in a structure (a quick sketch or the scene's own edges) for placement. Choose a palette target from your composite.
Step 2: Generate into a selection · 10 min
- Select the area for the new element in your composite. Generate with a short prompt plus the reference/structure, so it fills your hole using the scene's context.
Step 3: Curate · 8 min
- Audition the variations; keep the one that fits composition and light best. Keep it on its own generative layer — don't flatten.
Step 4: Integrate & finish · 10 min
- Apply Module 2 integration — edge, grain/CA, depth-of-field — and hand-fix any specifics the model got wrong. One unifying grade over all. 🏆
✅ Project Completion Checklist
- ☐ Generation directed by a reference + structure, not a bare prompt
- ☐ Generated into a selection so it used the scene's context
- ☐ Best variation curated; kept on its own layer
- ☐ Integrated (edge, grain, DoF) and specifics hand-fixed
- ☐ One unifying grade; result survives a 100% audit
🧗 Now You: Solo Variation
🌟 Your challenge
- Generate the same element with a bare prompt and again with reference+structure; count the re-rolls each needed. The gap is the lesson.
- Use a structure sketch to force a specific composition, then swap style references to reskin it three ways without moving anything.
- Generate purely as concepting — a dozen thumbnails to explore a direction — then rebuild the chosen one properly by hand. AI as ideation, craft as execution.
Going further: feed a matte-painting block-in (Module 4) as your structure so the model fills your composition with texture, then overpaint. AI plus matte painting is a potent, fast concept pipeline.
🍳 Recipe Card: Pro Generative
Control, then finish
- Pin composition (structure), look (reference), palette (style ref)
- Keep the prompt short; let constraints steer
- Generate into a selection so it uses the scene's context
- Curate the best variation; keep it on its own layer
- Integrate (Module 2) + hand-fix specifics + one unifying grade
Mantra: don't prompt-and-pray — constrain, then curate, then finish.
📓 Learning Journal
Add to your journal after this lesson:
- Key concepts you learned
- Techniques that clicked for you
- Questions or confusion points to revisit
- Ideas you want to try
- Your progress and feelings about learning this
✍️ This lesson's prompt: Where does generative AI genuinely fit your workflow — concepting, filling, extending — and where does it cost more to fix than to make by hand? Draw your own line.
🔭 Where This Leads
You can direct generation from scratch; next you use AI as an assistant on existing images. Lesson 7.2 — Neural Filters & Hybrid AI+Hand Workflows folds targeted AI tools into a hand-led process where you stay firmly in charge.
Is using AI "real" art or design?
The judgment, direction, integration, and finishing are yours, and they're where the value lives — the model is a tool like any brush or filter. The debate is real, but professionally the question is simpler: did you take responsibility for the result and its rights? That's Lesson 7.3.
Why not just re-roll the prompt until it's perfect?
Because it wastes time, credits, and rarely converges on your composition. Constraints get you there faster and give you a result you can defend and reproduce. Re-rolling is gambling; directing is designing.