Image Generation for Previz

For most of my career, "show them the film before you shoot it" was a rich production's privilege. Concept art meant hiring an illustrator. Storyboards meant hiring another one, or drawing stick figures and apologizing. Lookbooks meant hours ripping frames from other people's films and hoping the reference did the talking. Image generation flattened that overnight: the previsualization layer of filmmaking — every job whose product is a picture of the film that doesn't exist yet — is now available to anyone who can describe a shot. That's not hype. It's the most immediately useful thing in this course after the writing tools.

And describing a shot is the actual skill. A director who knows lenses, lighting, and composition gets dramatically better results from Midjourney than a civilian, because the prompt is just shot language: 35mm, low angle, sodium-vapor streetlight, fog, negative space left of frame. You already speak the language. The generator is the first illustrator who's fluent in it too.

The previz assets to generate

• Concept frames — single images that nail the mood of a key scene. The fastest way to find out if the film in your head survives being looked at.
• Storyboards — rough panels for coverage planning. They don't need beauty; they need camera position, staging, and eyeline, one panel per setup.
• Lookbooks — a coherent set of frames defining palette, wardrobe, texture, and light, so your DP and designer argue with pictures instead of adjectives.
• Location & design studies — variations on a set, a costume, a creature, a prop, iterated in minutes before anyone spends real money building or scouting.
• Pitch-deck frames — the mood pages of your deck, generated in your film's own look instead of borrowed from someone else's.

The craft problem you'll hit immediately is consistency — generators love giving you a new protagonist and a new house in every frame. Three habits fix most of it. First, work from reference images: feed the tool your location photos, wardrobe tests, or one approved character frame, and generate variations from that anchor rather than from scratch. Second, build a style prompt and freeze it — one saved block of look language (stock, palette, lighting, era) pasted into every generation, so the lookbook reads as one film instead of forty. Third, iterate a single frame until it's exactly right, then use that frame as the seed for everything else. Boards made this way won't be perfect — hands will be weird, geography will drift — and it doesn't matter, because boards are a communication tool, not a deliverable.

The rule, and where the payoff lands first

That pullquote is the rule, so let me give its reasons. The rights problem: as we covered in Chapter 4, purely AI-generated images generally can't be copyrighted, and the training-data litigation is unresolved — putting generated frames on screen imports both problems into your finished film, where a distributor's lawyer will eventually find them. The taste problem: generated imagery has a look, that over-lit, over-detailed, faintly airbrushed sheen, and audiences are already learning to clock it. In a pitch deck it reads as vision on a budget. Cut into your film, it reads as a shortcut. Keep the generator on the planning side of the camera and both problems stay out of your movie.

And here's where it pays off first: the pitch deck. Funders — investors, grant panels, that dentist who might come in for a few thousand — mostly can't read a script the way you can. They've been imagining a beige version of your film, if they're imagining anything at all. Twelve frames in your actual intended look changes the entire conversation, because for the first time they can see the film — your lighting, your palette, your faces. Nothing else in this course closes the gap between "trust me" and "look" as cheaply.

Previz is images that stand still. Next chapter, we get to the loudest tools in the shed — Sora, Veo, Runway — and what AI video generation is actually good for on a real production right now.