7. Chaining Shots Into a Finished Scene

Handing one shot to the next

Your Shotcut timeline now holds nine shots of the courier scene. Six are moving pictures: the two clips you started with, the crouch and the rise, plus the four new shots from the last chapter. Those four are the phone close-up, the sign insert, the reaction close-up and the wide walk toward screen right. The other three are still stills. They hold their place and their length, and they move nothing.

This chapter turns those three stills into shots. Then you watch all nine as one cut. The question is simple to say. How do you make a new shot pick up exactly where the last one left off? Runway can generate each shot well on its own. The hard part is the join between two shots, because a viewer's eye crosses it in a fraction of a second.

As before, the walkthrough is illustrative. The shots, the takes and the credit numbers are examples to show the method. They are not results anyone measured, and the series showing a workflow does not mean you have run it.

A cut is a hand-off

Think about what a cut does. One shot ends. The next begins. If the action keeps going across that cut, the viewer expects the person to be in the same place, in the same pose, moving the same way. Her arm is still rising. Her foot is still mid-step. The rain is still falling from the same side.

Sometimes a cut changes the framing instead. You jump from a wide shot to a close-up of something new. The viewer accepts a fresh picture there. What they still expect is the same person, the same clothes and the same light.

These two kinds of cut want two different starting frames.

When the action runs straight across the cut, the best first frame for the new shot is the last frame of the old one. You take the final picture of the accepted clip, save it as a still, and give it to Gen-4.5 Image to Video as the new shot's first frame. The new shot then starts from a pose that already matches. This is frame chaining. Each shot hands its last frame to the next, like a runner passing a baton.

When the framing changes, a chained frame does not help. The last frame of a wide shot is a wide picture, and the next shot is a close-up. You would only be asking the video model to zoom and invent detail. There, you go back to what you did in the last chapter. You make a fresh first frame in Gen-4 References from your tagged reference images, and you animate that.

There is a third case, and you already know its tool. Sometimes the new shot has to end on a pose that matches a shot you already have. Then both ends matter. That is the job of Runway's Animate Frames app, which you used in chapter four to steer a shot from both ends.

Reading the three stills against the shot list

Open the shot list next to the timeline. For each still, read its row: framing, action, screen direction, eyeline and length. Then read the scene beats around it. You are asking one thing about each cut on either side of the still. Does the action keep going, or does the picture change?

Say your three stills are these. First, the knock at the bakery door, which opens the scene before the existing crouch clip. Second, the courier reaching the door under the lit window next door, right after the wide walk. Third, the insert of the name plate on that door, the detail your animatic showed was missing.

Now take them one at a time.

The knock comes before the crouch. The crouch clip already exists and is accepted. You will not regenerate it. So the knock must end in a pose that flows into the crouch clip's first frame. Its start is a fresh picture, since it opens the scene. Its end is fixed by a shot you already own. Both ends matter, so this one goes to Animate Frames.

The arrival at the lit door comes straight after the wide walk toward screen right. She is walking in one shot and still walking in the next, closing the last few steps to the door. The action is continuous. This is the chain.

The name plate insert is a close-up of a sign on a door. The framing changes completely from the shot before. A chained frame would be a person at a door, not a name plate. This one gets a fresh Gen-4 References first frame.

So the three remaining shots use three different starting methods. That is the point of reading the list first. Each method costs credits, and choosing by the cut keeps you from spending them on the wrong one.

Exporting the last frame

Start with the chain, because it is the new skill. You need the final frame of the accepted wide walk, saved as a still.

You can do this in Shotcut. Open the accepted wide walk clip in the Source viewer, or find it on the timeline. Move the playhead to its very last frame. Use the arrow keys to step one frame at a time. Left and right move you a single frame. Step to the end, then check that the picture is really the last frame of the shot. Some files have an empty or held frame at the tail. You want the last frame that shows real motion, not a blank.

Then go to the File menu and choose Export Frame. The shortcut is Control, Shift and E on Windows or Linux, and Shift, Command and E on a Mac. Give the file a clear name, such as the shot number with the words last frame, and save it as a PNG. A PNG keeps the picture without the extra blur that heavy compression can add. Shotcut saves the still at your project's resolution.

Runway can also hand a frame forward without leaving the browser. In its video viewer you can scrub the clip to a frame and use that frame as the image input for a new generation. That is quicker. Shotcut has one advantage here. You are looking at the clip exactly as it sits in your cut, at your project's size, and you keep a named file in the project folder. That file becomes part of the record of how the shot was made. For this chapter, use Shotcut and save the still.

Checking that the frame is clean

Before you upload anything, look hard at the still. Zoom in on her face, her hands and the edge of the bag.

The thing to check is blur. The last frame of a walking shot often catches a foot or a hand mid-swing. In a moving clip, that blur reads as speed and looks natural. Frozen as a still, it is a smear. If you give a smeared frame to the video model, it treats the smear as part of the picture. The blur can come through into the new shot as a soft hand or a melted sleeve.

If the last frame is smeared, step back one or two frames and export again. Find a frame where her hands and face are sharp. You lose a tiny sliver of time at the end of the walk, which you can trim in Shotcut later. Then trim the end of the wide walk on the timeline to finish on that same frame. The cut will then join on the exact picture the next shot starts from.

Check the size and shape too. Gen-4.5 Image to Video takes PNG, JPEG or WebP images. The video it makes follows the shape of the image you give it. Your project is landscape, sixteen by nine, so the still should be too. If the image does not match the generation setting, Runway offers to crop it, and a crop changes the framing. Keeping the same project size in Shotcut and in Runway avoids that.

Animating from the chained frame

Now open Gen-4.5 Image to Video. Upload the exported still as the first frame.

Write the prompt the way your production bible says. Video prompts describe motion and camera only. Do not describe the courier again. Do not repeat the jacket, the collar or the bag. The picture already holds all of that. When you describe those things in the video prompt, the model reads your words as a fresh request and may redraw them. That is how a collar flattens or a jacket changes shade between shots.

So a prompt for the arrival might be something like this, in your own words: she keeps walking toward screen right, slows over two steps, and stops facing the door, while the camera holds still at the same height. That is all motion and camera. Keep the camera words the same as the ones in the wide walk's prompt. If the wide walk used a fixed camera at eye level, say that again here. Matching camera words keep the two shots behaving like one piece of film.

Choose a short length. Five seconds is enough for a few steps and a stop. Short takes give the model less time to wander away from the first frame. The longer a take runs, the more the end of it can drift from the start.

Generate one take and bring it into Shotcut before you judge it.

Inspecting the new take

Put the new clip on the timeline right after the wide walk, replacing the still. Then play across the cut, not just the new shot by itself. A shot can look fine alone and fail at its join.

Check these things in order, and look at the picture rather than trusting the prompt.

Identity first. Is this the same face as the sheet? Same hair, same jaw, same age.

Then the wardrobe details from your crops. Is the collar up, as the bible says? The collar is the detail that failed last chapter, so give it a slow look. Is the bag on the same shoulder as in the wide walk?

Then light. Does it fall from the same side as in the wide walk? Light that suddenly swaps sides makes a cut feel wrong, even when the viewer cannot say why.

Then screen direction. She was walking toward screen right. In the new shot she must still be heading right. If she turns and walks left, the viewer will feel she has changed her mind.

Then eyeline. When she stops at the door, where is she looking? The shot list says what she looks at next. Her eyes should point there.

And one more check, at the join itself. Image to Video starts from a frozen picture. It does not know how fast she was moving in the shot before. So a chained shot can start with a small lurch, as if she sped up, or with a pause, as if she stopped and started again. Play the cut a few times at normal speed. If there is a little settle at the start, you can often trim the first few frames of the new shot in Shotcut so the movement flows.

Say your first take fails one check. She arrives at the door, but the bag has slid to the other shoulder halfway through. That is a rejected take. Log it. Its credits count toward the shot, even though it will never be seen.

Ask what caused it before you regenerate. The first frame was correct, so the picture was not the problem. Look at your prompt. Say you had written that she adjusts her bag as she slows. The model took that as a reason to move the bag. Remove that action. Each shot has one action, and this one is arriving at the door. The second take keeps the bag where it was. It passes every check. Accept it.

When a chain drifts

You have chained one link. Now picture a longer chain.

Say a later scene needs the courier to walk a whole street in four connected shots. You export the last frame of shot one and animate shot two from it. Then you export the last frame of shot two to start shot three. And so on.

Each link copies the frame before it, and each copy is a little less faithful. The model rebuilds the picture every time. Small changes stack up. By the fourth link the face can be slightly different. Fine details soften. The jacket's grey might warm. The street behind her may bend. None of these changes is large between two neighbouring shots. That is why it is easy to miss. You only see it when you compare the last link to the first, or to the reference sheet.

So the check for a long chain is not just shot to shot. Put the last link's frame beside the reference sheet's three-quarter view. Then put it beside the first frame of the chain. If the face, collar or colour has moved, the chain has drifted.

The fix is to re-anchor. Instead of exporting yet another last frame and passing the drift along, go back to the source. Make a new first frame in Gen-4 References from the reference sheet and your crops, using the bible's fixed description line. Give it the pose and framing the chain had reached. Then animate from that clean frame. The chain carries on from a picture built from the sheet, not from a copy of a copy.

Re-anchoring costs a little more for that link, since you pay for the References still. It saves far more than it costs, because it stops the drift before it spreads to every shot that follows.

Some habits slow drift in the first place. Keep each link short. Use the same camera and style words in every link's prompt. If your Runway workflow lets you reuse a seed, keeping the seed from the first good take can help the takes behave alike. A seed only helps as far as the tool allows, though, and it does not guarantee the same result. Chaining also gives no discount. Each link is billed like any other generation, so a long chain is not cheaper than separate shots.

When the end pose must match too

Re-anchoring fixes the start of a shot. Sometimes the end matters just as much. That is your knock shot.

The knock opens the scene, and it must flow into the crouch clip you already accepted. So you need two stills. The first is a fresh Gen-4 References frame: the courier at the bakery door, hand raised to knock, built from the sheet and the bible's fixed line. The second is the crouch clip's first frame. Export it from Shotcut the same way you exported the last frame of the wide walk, with Export Frame, after checking it is sharp.

Now open Animate Frames, under Apps and Workflows in Runway. As in chapter four, it needs a paid plan, and both stills must share the same shape. Load the knock frame as the start and the crouch frame as the end. The prompt again describes motion, timing and camera only: she knocks twice, waits, lowers her arm and begins to crouch, while the camera holds still. Choose five seconds. The output is silent, which is fine, since the scene's ambience bed already runs under every shot.

Because the end frame is the crouch clip's own first frame, the knock shot lands exactly on it. When you place the two clips side by side, the cut joins on matching pictures.

Inspect it like the others: face, collar, bag side, light, eyeline. Her eyes should be on the door while she waits. Say this take passes. Accept it.

The name plate insert

The last still is the easy one, because you have done it before. The name plate is a close-up of a door sign, and it is a new framing. Make a fresh first frame in Gen-4 References. Tag the location reference and, if her hand appears reaching for the bell, the courier sheet. Then animate it in Gen-4.5 Image to Video with a short, quiet move, such as a slow push toward the plate. Check it and place it.

All seven new shots are now moving pictures. Every still is gone from the timeline.

Watching the scene as one cut

Now watch the whole thing. Play all nine shots from the knock to the name plate without stopping. Do not pause on any one shot. Watch it the way a viewer will.

Then watch it with the sound off and ask the scene questions from chapter five. What does she want? What blocks her? What changed? If a viewer can answer all three from the picture alone, the scene works as a story.

Then watch it a third time for continuity only. Look at each cut. Does the collar hold from shot to shot? Does the bag stay on its shoulder? Does the light stay on its side? Does she keep heading screen right once she sets off?

Say this pass shows one fault. The name plate insert looks lit from the left, while every shot around it has the window glow coming from the right. Alone, the insert looked fine. In the cut, it jumps.

Trace it back, the way you traced the collar last chapter. The video prompt was only motion and camera, so it did not set the light. The first frame did. Open the first-frame prompt. Say you find that the bible's fixed lighting words were left off, so the model chose its own light. That is the single input at fault.

Fix only that. Rewrite the first-frame prompt with the bible's lighting words. Make a new References frame. Check it beside the shot before it. Animate it with the same motion prompt as before. One new take. Place it and watch the cut again. The light now holds.

Every other shot stays exactly as it was. That is the milestone. The scene's continuity survived a repair to one shot without starting over, because each shot's inputs were saved and the fault could be traced to one of them.

Logging the cost

Open your cost log and add the new work. These numbers are illustrative.

Say the knock cost about fifty credits for the Animate Frames take, plus about eight for its References frame, so about fifty-eight. The arrival needed two takes of about sixty each, one rejected and one accepted, so about a hundred and twenty. It had no References cost, because its first frame came from your own clip. The name plate cost about sixty-eight the first time, frame and take together. Its repair cost about another sixty-eight.

Add those and the three new shots come to about three hundred and fourteen credits. The log already held about two hundred and fifty-two for the first four shots. Together that is about five hundred and sixty-six credits for seven accepted shots, or about eighty per accepted shot.

Your planning estimate in chapter five was about sixty-six per shot. The real figure came in higher, mostly because of one rejected take and one repair. That gap is the reason to log rejected work. Generated seconds are not finished footage. The next time you estimate a scene, you can plan with eighty, not sixty-six.

Save the exported last frame, the knock's two stills, the rejected takes and the new prompts in the project folder beside the rest. Each shot's record should let you rebuild it.

The picture is locked; the sound is not

The picture is now locked. Nine shots tell the scene. Play it once more and listen. Under the whole scene runs the ambience bed you added in chapter two, and nothing else. Her knock makes no sound. Her steps make no sound. The parcel lands on the step in silence. Animate Frames gave you silent shots, and Image to Video added none on top of your bed.

A locked picture is the right point to start on sound. The timing of every shot is fixed now, so each knock, each step and each set-down has a place to land. That is the next stage of the scene.

Kling 4.0 and the long hand-off

One recent release bears directly on what you just did. Kling AI has launched Kling 4.0 on its web platform. Three of its changes aim at the same problem as frame chaining: keeping a character steady across a hand-off.

First, it extends how long a single pass can run. Your arrival needed two separate shots joined by a chained frame. A longer pass could, in principle, carry the wide walk and the arrival as one continuous take, with no join to match at all. Whether it holds the courier steady across that length is something to test, not assume.

Second, it accepts several keyframe images placed along the timeline, not just a start and an end. Animate Frames lets you fix two points. Kling 4.0 lets you fix several, such as a pose at the start, one mid-walk and one at the door.

Third, its Omni Reference system combines many reference assets at once, including multi-angle character sheets and reference videos. Your three-view sheet and your crops are exactly that kind of material.

Some things are not there yet. Developer access through an API, ten-bit high dynamic range output and multi-minute extensions are listed as coming soon. There is a free tier with daily credits, but its output is watermarked and for personal use. Paid plans give commercial rights and watermark-free exports. Monthly subscription credits expire without rolling over, while bought credit packs last for years. Higher resolution costs more credits per second, with ten-eighty above seven-twenty.

The smallest useful next step is one test on a shot you already have. On the free tier, upload your three-view courier sheet and the collar and bag crops. Give it the first frame of your wide walk and the last frame of your accepted arrival as keyframes, and ask for one continuous take. Then put it in Shotcut beside your two chained shots and inspect it the same way: face, collar, bag side, light, screen right. Remember the watermark makes this a test, not footage for the finished short. If the single pass holds the courier better than your chain did, you have a reason to price it properly. If it drifts, your chain stays.