6. Keeping Mara the Same Person

Writing continuity down once

Seven new shots stand between the animatic and a real scene. Each one will be generated separately. Each one is a fresh chance for Mara to come back slightly different: a darker coat, a bag on the wrong shoulder, light from the wrong side of the street. This chapter answers one question. How do you make seven separate generations show the same woman on the same evening?

The answer is a habit more than a trick. Continuity is a set of decisions you write down once and feed into every shot. If you rely on remembering what Mara looked like each time you type a prompt, she will drift. So the work happens in three parts. First comes a reference sheet, a set of clean pictures of Mara from several angles. Next comes a production bible, one page of written rules for the scene. Last comes generation, where the pictures and the rules go into every shot. Then we break one shot on purpose and repair only that one, to prove the system holds.

Everything here builds on what is already in the project folder. That includes the accepted Mara reference still and its revision, the crouch and rise clips, the bakery-step still, the nine-shot list and the Shotcut animatic. Everything below is an illustrative walkthrough, and the numbers in it are examples. It is not a record of anything you have done.

The tool and what it costs

The hosted tool stays Runway, which the course has used since the first chapter. For this job Runway splits the work into two stages, and the split matters.

The first stage is Gen-4 References. It is an image generator, not a video generator. You give it one to three reference images and give each one a tag, a short name that starts with an at sign, such as at-mara. Then you write a prompt that uses those tags. Gen-4 References makes a new still that combines them, so Mara from your reference stands in a setting or pose your prompt describes. It accepts PNG, JPEG and WEBP files. Each image costs about eight credits.

The second stage is the image-to-video method from chapter three. The still you made with References becomes the first frame of a Gen-4.5 Image to Video generation. As chapter three showed, Gen-4.5 runs anywhere from two to ten seconds, so a shot can be generated at the length the shot list asks for. Its base output is 720p-class, which means twelve hundred and eighty by seven hundred and twenty pixels at sixteen by nine. Larger sizes need an upscale. Gen-4.5 Image to Video gives silent output, and that suits this scene. The ambience track in Shotcut already carries the sound.

Why two stages, and not one? The video model cannot take a loose pile of reference pictures and work them all in. It anchors on one first frame. So identity gets locked into a single still first, and the motion comes second. That is why the reference sheet matters so much. It feeds the still, and the still decides the shot.

On plans: Runway's free tier gives a one-time grant of a hundred and twenty-five credits that does not refill, and its output is watermarked. The Standard plan gives six hundred and twenty-five credits a month, and paid plans carry commercial licensing. Pro gives two thousand two hundred and fifty. The old estimate for the seven new shots was about four hundred and sixty credits. That fits inside one month of Standard, but not by much.

Building the reference sheet

A reference sheet is a set of pictures of one character from fixed angles, under plain light, on a plain background. Animators have used them for decades so that many artists draw the same face. Here the artist is a model, and the sheet does the same job.

The sheet needs three views. The front view shows Mara looking straight at the camera. The three-quarter view has her turned about halfway between front and side. The profile shows her exactly side-on. These three cover the angles the shot list uses. The phone close-up sees her face from slightly above and to the side. The reaction close-up sees it nearly front-on. The walk away sees her back and side.

Start in Gen-4 References. Upload the accepted Mara reference still and tag it at-mara. Then prompt for the front view: "at-mara, front view, full body, standing straight, arms relaxed, plain mid-grey background, flat even studio light, no shadows." Generate it. Change only the angle words and make the three-quarter view, then the profile. That comes to three images at about eight credits each, so roughly twenty-four credits.

Why flat light and a grey background? Because the model copies whatever it sees. Suppose your reference shows Mara lit warm from the left in front of a brick wall. Then the warm light and the bricks become part of what the model thinks "Mara" is. It may carry them into a shot where the light should come from the right. Flat light shows the true colour of her coat and the true shape of her face, with no shadows hiding the details. The light of each shot then comes from the shot's own prompt and its own setting image, and from nowhere else.

Your first try will probably not be perfect, so check each view before you keep it. Pull two frames from the clips you already accepted: one from the crouch, one from the rise. In Shotcut, park the playhead on a clear frame and use Export Frame from the File menu. Now set the new views beside these frames and compare them one feature at a time. Is the hair the same length, and parted on the same side? Is the coat the same colour, or has it shifted toward brown or grey? Is the collar shaped the same? Is the bag strap on the same shoulder?

In this illustrative pass, the profile comes back with the hair tied up, while the clips show it loose. That profile is rejected. You log it as eight credits spent, with a note saying hair tied. Then you regenerate with "hair loose to the collar" added. One rejected image is normal. Keeping a wrong view is much worse, because every later shot would inherit the mistake.

Then come the detail crops. A full-body view is too small to show the things a close-up will show. Open the best front view in any image editor and cut out three close crops. The first is the collar: its shape, its colour and how it sits at the neck. The second is the bag: its strap, its buckle and which side it hangs on. The third is the parcel label: where it sits on the box and roughly what it looks like. Save each crop as a PNG in the reference folder. The crops cost nothing. They are just closer looks at pictures you already have. They earn their place later, because References takes only three images per call. When a shot is a close-up of the collar, the collar crop is a better choice than a full-body view where the collar is a few pixels wide.

The one-page production bible

The sheet holds what Mara looks like. The bible holds everything else the scene must keep constant, written in words so you can copy it into prompts and check shots against it. Keep it to one page. A page you actually read is better than five you skip.

Write it in six short parts.

Appearance is a sentence or two about her face, hair and build. In this illustrative bible: late twenties, dark hair loose to the collar, parted on the left.

Wardrobe means the exact colours and fabrics. Vague words let the model wander, and exact ones hold it. "Dark clothes" could mean anything. "Matte navy canvas courier jacket, collar up, zip closed to the chest" means one thing. The same goes for her trousers and shoes.

Props are the bag, the parcel and the phone. Say which shoulder carries the bag strap. In the illustrative project it is her right. Describe the parcel's size and the label's position, and the colour of the phone case.

Lighting is where the light comes from on the bakery street and what time it is. Here it is early evening, a warm streetlamp key from screen left, and a cool blue sky fill. Use fixed technical words and reuse them exactly every time. Mood words like "moody" or "cosy" mean something slightly different to the model each time. "Warm streetlamp key from screen left, cool dusk fill" means the same thing every time.

Screen direction says that Mara exits toward screen right, as the shot list already says. Any shot that shows her walking must keep her moving right. If one shot flips her to walking left, the viewer will feel she turned around.

The fixed description line is one sentence that names Mara the same way every time. For example: "Mara, a courier in a matte navy canvas jacket with the collar up, dark hair loose to the collar, bag strap on her right shoulder." You paste this sentence word for word, every time.

Where does that line go? This is where the two stages matter. Runway's guidance for image-to-video is to describe only motion and camera in the video prompt, because the first frame already sets the clothes and the face. If you describe the wardrobe again in the motion prompt, the model may change the clothes when it should keep them. So the fixed line goes into every Gen-4 References prompt that makes a first frame. The Gen-4.5 video prompt gets only what moves and how the camera behaves. The line is fixed in its words and fixed in its place.

Each part of the bible stops a particular kind of drift. Appearance and the sheet stop a new face. Wardrobe words and the collar crop stop the jacket changing. Props stop the bag jumping shoulders. Lighting stops a cut where the sun seems to move. Screen direction stops a reversed walk. The fixed line stops the small rewordings that creep in when you type a description fresh each time.

Generating the new shots

Four of the seven new shots go first: the phone close-up, the sign insert, the reaction close-up after the sign, and the wide shot of Mara walking right toward the lit flat. Choose them because they are the most different from each other. One is a close angle looking down, one is a straight-on face, and one is a full figure in motion. If the system holds across those, it will probably hold across the rest.

Each shot follows the same steps.

1. In Gen-4 References, choose up to three references for this shot: usually one view of Mara, one detail crop and the bakery-step still for the street.
2. Write the first-frame prompt with the tags, the fixed description line pasted unchanged, the framing from the shot list and the lighting words from the bible.
3. Check the still against the bible before you spend anything on video.
4. Use that still as the first frame in Gen-4.5 Image to Video, set the shot list's duration and write a prompt describing only the motion and the camera.
5. Inspect the clip, then log it as accepted or rejected with its credit cost.

Take the phone close-up as the worked example. The shot list says close-up, Mara looks down at her phone, eyeline down, three seconds. The references are at-mara from the three-quarter view, at-bag from the bag crop because the strap will be in frame at the edge, and at-street from the bakery-step still. The prompt reads: "at-mara on at-street, close-up from slightly above," then the fixed description line, then "looking down at a phone held in her left hand, warm streetlamp key from screen left, cool dusk fill." The still comes back. Check it against the bible before going further. Jacket navy, collar up, strap on the right, light from the left. It passes, at about eight credits.

In Gen-4.5, that still is the first frame. Set the duration to three seconds. The prompt reads: "she scrolls with her thumb and her eyes move across the screen, camera static." Nothing about the jacket, because the frame has already decided it.

Now inspect what happened, not what you asked for. Play the clip several times and watch one thing each time. First the face: does it stay the same face as the sheet from first frame to last, or does it soften or shift halfway? Then the collar: does it stay up? Then the bag strap: does it stay on her right shoulder, or slide? Then the light: does the warm side of her face stay on the left the whole way? Then the eyeline, which the shot list says is down toward the phone. If her eyes drift up toward the camera, the shot will not cut against the next one. In this illustrative pass the phone close-up passes all five checks.

The sign insert is simpler, because it is a shot of the "closed for the week" sign with Mara out of frame. Its references are the street still and nothing else. One thing to watch: generated text is often scrambled. Read the sign letter by letter. If it is unreadable, a real fix is to make the sign as a clean image yourself, add it in the edit, and generate only the shopfront. A model that keeps getting text wrong is not worth more spends.

The reaction close-up after the sign is the shot the animatic showed was missing. Its references are at-mara from the front view, at-collar from the collar crop because the collar fills the bottom of the frame, and at-street. Its eyeline is up and to the left, toward the sign. Check that eyeline carefully, because the cut from the sign to her face depends on it. If she looks up and right, the viewer will think she is looking at something else.

The wide walk uses the profile view, the bag crop and the street. The motion prompt says she walks toward screen right, and screen direction is the extra check here. Watch the whole clip and make sure she never turns back.

Judging the cut as a sequence

Four shots that each pass inspection still might not work together. You can only tell by watching them as a sequence.

Open the Shotcut project with the animatic. Import the new clips into the playlist. For each, find the still it replaces on the timeline. Drag the new clip onto the track above that still, starting at the same point. Trim it to the still's length, then delete the still underneath, or mute that lower track, so the timing the animatic proved stays the same. Do this for all four. The two clips you already had, the crouch and the rise, stay where they are. The three shots you have not generated yet stay as stills for now. A mixed timeline is normal at this stage.

Then play the sequence from the start, at full speed, without stopping. You are no longer checking single frames. You are checking the cuts. Does the jacket look like the same jacket across every cut, or does one shot jump slightly brighter? Does the warm light stay on the same side as the shots change? When the sign insert cuts to her reaction, does her look seem to land on the sign? When she walks off in the wide, does her direction match the rise clip from chapter four?

In this illustrative pass, three of the four sit together well. The reaction close-up does not. Watched alone it looked fine. Watched after the phone close-up, the collar is plainly wrong. In the phone shot it stands up. In the reaction shot it lies flat, and the jacket is open lower at the chest. A viewer might not name the problem, but they would feel Mara had changed.

One repair without starting over

This is where the system earns its keep. Chapter three showed that simply regenerating a wrong shot proves nothing about what caused the problem. Now there are written decisions and fixed references, so you can ask exactly where the error came in.

Work backward. The clip started from a first frame. Open that still. The collar is already flat there. So the video stage did not cause the problem. It faithfully animated a frame that was wrong from the start. Running Gen-4.5 again would waste credits, because it would start from the same bad frame.

Now look at the still's inputs. The references were the front view, the collar crop and the street. The fixed line said "collar up." So why flat? Compare the collar crop with the front view, and the answer shows. You cut the crop from a front view where the collar sat a little low. The crop showed a soft, half-folded collar. The model followed the picture over the words.

The repair is aimed only at that cause. Make a new collar crop from the three-quarter view, where the collar clearly stands up. Swap it in for the old crop in this one shot's references. Change nothing else: not the prompt, not the street, not the other three shots. Generate the still again, about eight credits. Check it. Collar up, zip at the chest. Then run Gen-4.5 from the new frame with the same motion prompt and the same three seconds. Inspect the new clip with all five checks. Put it on the timeline in place of the rejected one and play the whole sequence again. The cut from the phone close-up to the reaction now holds.

Two things made that repair small. First, every input was saved and named, so you could find exactly what went into the bad shot. Second, the other shots never depended on the bad crop. Each shot reads from the reference folder, not from the shot before it. So fixing one shot leaves the others alone. That is the milestone for this stage of the course: a multi-shot scene whose continuity survives a targeted revision without starting over.

The bible gets a one-line update too: "Collar crop: use the three-quarter view crop; the front-view crop reads as a soft collar." Your next shot then starts from the corrected reference. The fix is now part of the decisions you wrote down.

Counting what it actually cost

Now the log. Every generation goes in it, accepted or not, with its cost read from Runway's generation details. The per-take video cost below is illustrative, because your account's figure is the one that counts. Suppose each video take is three seconds, like the phone close-up and the reaction close-up. At Gen-4.5's twelve credits per second, that is about thirty-six credits a take.

On the image side there were three views plus one rejected profile, which is four stills. There were four first frames for the shots, one of them the rejected reaction frame, plus its replacement. Nine images at about eight credits each is roughly seventy-two credits. On the video side there were four first takes plus one repeat of the reaction close-up, so five takes at thirty-six is one hundred and eighty. The total is about two hundred and fifty-two credits for four accepted shots, or about sixty-three per accepted shot.

That is just under the sixty-six per shot the animatic estimate allowed. The log shows where the money went. Nearly half the image spend went on the sheet, which is a one-time cost. The remaining three shots will reuse it and pay nothing for it. One rejected video take, though, cost more than all the rejected stills together. That is the lesson the numbers teach. Check the cheap still before you pay for the expensive motion, every time. Each eight-credit check that catches a mistake saves a thirty-six-credit take.

The sound is still only the ambience bed, and Mara has no lines. Sound design, dialogue and finishing remain untaught. Three new shots are still stills on the timeline, and they are next in the queue with their references already chosen. One problem this chapter's method does not solve is the moment where one shot's last frame must hand off exactly to the next shot's first, as when she rises and turns. For those joins the next tool is frame chaining, where the end of one accepted clip becomes the start of the next.

Before that, one release this month is worth a look, because it changes how many references a single generation can hold.

ShengShu Technology released a preview of Vidu Q4 on the seventh of October. In one generation setup it accepts up to fifteen reference images and three reference audio files. Compare that with the three images Gen-4 References takes per call. In the repair you just did, the three-image limit forced a choice: a body view, a detail crop and the street, and nothing more. With fifteen, the whole reference sheet could go in at once. That means all three views, all three detail crops, the street and a frame from the accepted clips. ShengShu describes the images as locking separate elements, such as character identity, wardrobe, props and background, across a scene. The reference audio is meant to guide a voice's identity and tone across shots, which matters later when Mara speaks. Output is native 2K and 4K with ten-bit colour, which leaves more room when you grade later. Pricing starts at about one and a half cents per second.

What this changes is the production task you just did by hand: choosing which three references a shot gets. What it does not change is the method. A larger reference count still needs a clean sheet and a written bible. Fifteen badly lit references would teach the model fifteen versions of the wrong light. The sheet and the bible move to any tool unchanged.

The smallest useful next action is a single test, not a switch. Take the reaction close-up, the shot that failed, with your existing sheet and bible. Generate it once in Vidu Q4 with the full set of references. Then run the same five checks: face, collar, bag, light and eyeline. Drop it onto the Shotcut timeline in place of the accepted Runway version and watch the cut from the phone close-up. Log its cost in dollars beside your credit costs, so the two can be compared per accepted shot. One shot judged on your own scene tells you more than any launch reel. Since this is a preview, check its access terms and its licence for commercial use before you plan the rest of the scene around it.

Two other releases point the same way. HiDream's vivago R1 builds multi-shot sequences inside one conversation and keeps reference assets anchored across turns. Higgsfield now saves character attributes as reusable assets in a project library. Both are bets that keeping characters consistent should happen in one place, much like the reference folder you just built.