Runway Resources | What is AI previs? A guide to AI previsualization for modern production

Summary

AI previs uses generative image and video models to produce storyboards, concept frames and rough animatics from a script, shot list or reference images. It covers the exploratory half of previs faster and cheaper than a 3D pipeline, and on some productions it has already become the default: Runway's AI Media Report found an independent film studio building full-film animatics that are 80% generative video before a frame of live action is shot. It does not produce measurable space, so any shot with something physical to plan, meaning a rig, a performer or a set build, still needs a traditional previs or techvis pass.

What is AI previs?

AI previs is previsualization done with generative models instead of hand-drawn boards or manually built 3D scenes. It takes a script, a written shot list or reference photos and returns storyboard panels, concept frames and short motion clips in minutes, covering shot composition, rough camera movement, blocking and scene pacing.

If you already work with previs, the useful question is not what AI previs is but which parts of the process it can absorb. AI previs can take on over the half of previs that exists to answer "does this scene work," and leaves the half that exists to tell a crew what to build. And today, previs itself has gone substantially generative on productions that still shoot.

A note on spelling: previs and previz are both short for previsualization. Previs is the form used by the Previsualization Society, the Visual Effects Society and most working crews, so this guide uses it throughout.

What AI can handle for previs

Previs does two jobs on a production. It communicates intent, and it produces a technical plan. Generative models are good at the first and have progress to make on the second.

Previs task	Can AI do it	Why
Storyboard panels from a script	Yes	Locked references produce usable panels in minutes once a person has made the shot breakdown.
Animatics and pacing tests	Yes	Panels on a timeline work the same regardless of how the panels were made. Generated clips hold looser timing.
Coverage exploration	Partly	Cheap enough to test whether a scene needs five angles, but the angles won't always share a consistent space.
Tone and lookdev	Yes	This is the strongest use, since the output is judged by eye.
Blocking against real geometry	Not alone	Models have no persistent 3D space to block within, but they will typically hold blocking rendered from geometry you built
Lens and camera data	Partly	Recorded metadata passes through cleanly and real footage solves to an estimate. Numbers pulled off a generated frame describe a camera that never existed.
Rig	No	Nothing models arm length, head clearance, tilt range or track length.
Continuity across a sequence	Partly	References hold characters and locations. Exact spatial continuity between shots does not survive.
Framing a generative VFX plate	Yes	Where an effect sits and what stays unoccluded is a framing question

The line runs through spatial grounding. A traditional previs scene is real 3D space with a real camera in it, which is why a lens length or a crane position can be pulled out of it and handed to a department. A generated frame is a picture of a camera move, not a camera move.

On the "yes" side, the compression is steep. Runway's AI Media Report, a longitudinal study across hundreds of enterprises with figures self-reported by customers against their prior-year spend on the same work, records an AAA game publisher cutting concept design from 30 days to one.

This is why AI previs does not replace a previs artist so much as change what they spend time on. Boards that were only ever going to be thrown away get made faster. The work that survives on a call sheet is unchanged.

How AI previs works

A typical workflow follows a consistent shape:

Break the script into scenes, beats and shots.
Build a shot list with intended framing, movement and emotional purpose for each shot.
Generate reference frames or storyboard panels from those shots.
Lock character, wardrobe, location and color references to keep the look consistent across shots.
Generate rough motion clips, or assemble static panels on a timeline as an animatic.
Review with the director, DP, production designer and editor.
Convert the approved ideas into a shooting plan, a techvis pass or final production assets.

Step four is where most attempts fail. A previs sequence is only useful if the same character, wardrobe and location stay recognizable across a dozen or a hundred frames, and prompt text alone will not hold them. Tools handle this by conditioning generation on locked reference images. In Runway that is References, which holds a character, location or style across shots; the equivalent exists under different names in most script-to-storyboard tools. Whatever the tool, lock references before generating volume, not after.

A director planning a chase sequence can build a 30-second animatic to test where the camera follows the action and when it cuts wide, before anyone scouts a location, and flag which beats will need practical stunts or VFX along the way.

AI previs vs. traditional previs

Traditional previs and AI previs both give teams something concrete to react to early. Traditional previs is built in real 3D space, so a crew can pull executable data out of it: lens choices, camera moves, set dimensions. AI previs trades that precision for speed, and in generative pipelines its output often becomes the shot rather than a reference for one.

Format	What it shows	Best for
Storyboard	Static key frames, hand-drawn or digital	Fast communication of shot ideas
Traditional previs	Rough moving 3D sequence, built in Unreal, Blender or a dedicated previs tool	Precise staging, timing and technically complex shots
AI previs	AI-generated images or video, guided by prompts or references	Early creative exploration and fast iteration

A previs artist building a sequence in Unreal or Blender controls camera lens, distance, blocking and timing down to the frame, which is why 3D previs remains the standard for stunt coordination, plate photography and anything that hands off directly to a technical department.

Generating a set of concept frames or a rough animatic takes minutes instead of days and requires no 3D modeling experience, which makes AI previs useful during pitches and early development when the goal is testing ideas rather than building a plan.

When previs output becomes the shot

On a live-action production, AI previs behaves like any other previs: the output is disposable. The animatic never appears in the film, the boards are reference material for the people executing the real shots, and the value is entirely in decisions made before the crew arrives. The AI Media Report cites an independent film studio building full-film animatics that are 80% generative video before a frame of live action is shot, which is a production still committed to shooting and simply doing its previs differently.

In a fully generative pipeline, there is no crew. Previs and production collapse into one loop, because the frame you approve becomes the frame a video model animates and the shot list becomes the generation queue. Previs output stops being references and starts being footage.

Tool's Under Armour campaign "Forever is Made Now" ran this way out of necessity, since the talent was unavailable for a shoot. The team generated stills in volume, curated them into groups and animated the selects, which meant the concepting pass produced the assets rather than a plan for them. Tool president Dustin Callif put the whole project at four weeks, one of them ideation and concepting, against a typical five to six weeks for a campaign with that many review rounds. The pipeline also drew on CG and licensed footage, so the compression is not attributable to previs alone.

The two paths have different standards. When previs output is disposable, rough is fine, and rough is often better because it keeps attention on staging rather than finish. When previs output becomes the shot, frame quality is footage quality, and a panel you approved at "good enough to read" ships at exactly that level. Teams moving between the two modes tend to get caught by this once.

Most productions run a mix. The exploratory sequences generate, the technically demanding sequences shoot, and the previs work splits accordingly.

Sometimes the previs pass is what gets a production made at all. History Channel's "Life After People" had its third season paused because a traditional VFX pipeline for sequences like the flooding of New Orleans would have needed a large team of senior artists working for weeks, and the budget was not available. Eggplant Picture & Sound was asked whether AI could finish the edits. Creative Director Geoff Marshall read the script, built a proof of concept and the show went back into production, eventually delivering 360 minutes across eight episodes with roughly a month of post time saved. The proof of concept did exactly what previs is for: it gave decision-makers something to react to before anyone committed a budget.

Marshall is also clear about the limit. Getting to a polish level where a viewer suspends disbelief, he says, requires the same attention to detail as traditional VFX. His team still built photorealistic matte paintings before generation and still composited practical elements over the output. Speed at the front of the pipeline does not remove craft further down it.

Where techvis fits

Techvis, short for technical visualization, is the precision layer between previs and the shoot. Previs answers whether a scene works creatively. Techvis answers what lens, rig, crane reach and camera position the shot requires, and whether the move is physically possible on the day. It is built in the same 3D tools as traditional previs but populated with real-world measurements: stage dimensions, set builds, motion control specs.

The standard sequence on a complex shot is previs, then techvis, then the shoot, then postvis, which combines real footage with temporary effects so an editor can cut before final VFX are done. AI previs sits upstream of all of it, and cannot substitute for the techvis step at any point.

Where AI previs is used

Early development and pitches. Producing a boarded sequence for a treatment or a client pitch, where the goal is showing tone and intent rather than handing off a plan. The binding constraint here is usually the deadline rather than the budget. A global consulting firm rebuilt a client campaign inside a two-day RFP window, which is not a timeline a 3D previs pass fits into.
Commercials and music videos. Short formats with compressed timelines, where testing five directions beats perfecting one.
Tone and pacing alignment. Getting a director, DP and production designer to agree on what a scene feels like before anyone is on set.
Coverage exploration. Testing whether a scene needs an extra angle, or whether two shots are doing the same job, at the point where cutting one costs nothing.
Productions without a previs budget. Most work has never had access to a previs artist or a 3D pipeline. For these teams, AI previs is not a faster version of an existing step; it is the first previs they have had.

When not to use AI previs

Hand off to traditional or technical previs in these cases:

Stunt sequences need real blocking against real physics, camera rigs and safety constraints, which is precision a 3D technical pass provides and a generative model cannot guarantee. Bodies and rigs do not negotiate.
VFX shot on a plate. Note the qualifier. The dividing line is not whether a shot is VFX, it is whether there is something physical to plan around. A creature comped into live-action footage needs the plate planned. A creature in a fully generated shot has no plate, no rig and nothing to measure, so techvis has nothing to contribute.

Generated visual effects are worth separating out, because they change what plate planning means rather than removing it. Generative visual effects workflows start from a still frame of live-action footage, generates the effect from that frame, then comps the result back into the original plate. There is also a more direct route, where a model like Aleph 2.0 edits the footage itself rather than generating a separate element to composite back. Either way, the planning requirements are real but different from techvis. A locked-off camera is preferred because it avoids complex tracking and masking, and the area where the effect will happen needs to stay unoccluded, or the shot picks up masking work later.

Both of those are previs decisions. Where the effect happens in frame, whether anything crosses in front of it, whether the camera moves at all: these get settled before the shoot. So a GVFX shot does not need less previs than a conventional VFX shot. It needs previs pointed at a different set of questions.

Shots that must match a real location's geometry need measurement-based previs, not generated approximation.
Anything going directly to a construction, rigging or stunt team needs frame-accurate camera data.

Generated video can also propose camera moves that look compelling but are not achievable with real equipment, or that break continuity from one shot to the next. Treat AI camera work as a statement of intent, not a rig plan, and confirm anything unusual with a camera operator before it reaches a call sheet.

The practical pattern: start with AI previs to find the story and the shape of a scene, then move the shots that need it into a traditional or technical previs pipeline.

Choosing an AI previs tool

The category splits by where a team sits in the pipeline. Tool descriptions and model names reflect each vendor's public materials as of publication and will date faster than the workflows around them.

If you want 3D precision on a tablet. Previs Pro (iPhone, iPad, Mac) builds actual 3D sets with characters, props and lighting, and includes an AR mode that places virtual characters into a real location through the device so you can frame shots by walking around them. It imports 3D scans as glTF or glb for accurate location backdrops, supports script import from Final Draft, Fountain, Celtx and open formats, and has a timeline for animatics with audio. Exports go to Premiere and Final Cut, plus MP4 and printable boards. It has added AI generation for characters and props, so it is no longer purely traditional.

If you want boards and nothing else. Drawstory parses a script, breaks scenes into shots and maintains character consistency, aimed at directors who want iteration without a 3D pipeline. It produces static panels, not motion or audio.

If you want boards through to a rough cut. M Studio is browser-based and carries storyboards into motion boards with audio. Script-to-storyboard output is editable shot by shot, character references stay locked across frames, and boards move into an animatic timeline with voice, music and sound effects in the same project. It runs on third-party generation models including Kling and Veo, and exports MP4, a timecoded review link and frame-by-frame PDF.

If you want the output to be footage. With Runway you can build storyboards, animate panels into a sequence, lay out a sequence from a single reference image, and use References to hold a character, location or style consistent across shots. Gen-4.5 generates and Aleph edits an existing clip, so an approved panel does not get rebuilt to become a shot. General-purpose video generators can produce individual clips, but a previs sequence needs the same cast and locations across dozens of them, and that consistency is the part they mostly leave to you.

Most film and VFX professionals treat AI-generated previs less like final art direction and more like fast composition testing: a way to check whether a scene idea works before investing artist time in fully realized boards.

Frequently asked questions

Can AI replace traditional previs?

Not entirely. AI previs covers the communication half of previs, which is showing a team what a scene looks like and how it paces. It cannot cover the technical half, because a generated frame contains no measurable space and nothing a rigging, construction or stunt team can execute against. On most productions AI previs runs first for exploration, then technically demanding shots move into a traditional previs or techvis pass.

Do I still need a previs artist if I use AI previs?

For technically demanding work, yes. A previs artist builds in measurable 3D space, which is what stunt coordination, plate photography and set-dependent shots require. What changes is the mix of their work: boards that were only ever going to be discarded get generated instead of built, so artist time concentrates on the sequences that have to be exact.

Does AI previs require 3D modeling skills?

No, and that is the main difference from traditional previs. Most AI previs tools work from a script, a written shot list or reference images, and generate output through prompts rather than manual scene building. This opens the workflow to writers, directors and marketers who never trained in 3D software. The tradeoff is that you give up the measurable 3D space a traditional previs tool produces.

How does AI previs save money on film production?

Mainly by moving discovery earlier. A problem that would otherwise reveal itself on a shoot day, such as an unclear stunt beat, a location that does not read the way the script implies or a VFX sequence nobody scoped, can show up during previs instead, when fixing it costs a few hours rather than a wasted call day. Runway's AI Media Report, which collects cost and time figures self-reported by enterprise customers against their prior-year spend, records a AAA game publisher cutting concept design from 30 days to one.

How early should a team start using AI previs?

As early as there is a script or even a rough outline. Generating a set of panels or an animatic takes minutes rather than days, so there is little cost to testing a scene's staging during a first pass rather than waiting for a locked draft.

Can AI do VFX, or only previs?

Both. Fully generated VFX shots have aired on national television, including season three of History Channel's "Life After People," where Eggplant Picture & Sound completed destruction and flooding sequences that had been shelved for budget. There is also a middle path, generative visual effects, where an effect is generated from a still frame of live-action footage and comped back into the original plate. That approach still requires planning the plate, but the requirements shift toward a locked-off camera and an unoccluded effect area rather than precise lens and rig data. Runway documents both on its AI for Visual Effects and GVFX pages.

Are there limits to AI-generated camera movement?

Yes. Generated video can propose camera moves that look compelling but are not achievable with real equipment, and it can break continuity between shots because the model has no persistent 3D space to move a camera through. Treat AI-generated camera work as a reference for intent, not a literal rig plan, and confirm anything unusual with a camera operator before it reaches a call sheet.

AI previs will not replace the parts of pre-production that depend on technical precision or hard-won craft. What it does is compress the distance between an idea and something a team can look at, so more ideas get tested and the ones that reach a set have already survived a few rounds of scrutiny.

A scene is easier to judge when a team can watch it instead of reading it. Build a boarded sequence with Runway and hold your cast and locations consistent across every frame with References, so the version your team reviews looks like the film you intend to shoot. Get started for free.