4. Steering a Shot From Both Ends

Planning the cut before you generate

The accepted crouch clip is three seconds long. Mara crouches and sets the parcel down on the bakery step, and the light holds steady from start to end. Watched alone, it works. A film is never one clip watched alone, though. The next thing the viewer sees has to follow from it. Mara straightens up. She turns toward the street. She walks away. This chapter builds that next shot, and it builds it with one question in mind. What will the cut between these…

Planning the cut before you generate

That question changes how you generate. Until now, each shot was judged on its own: is the face right, does the hand touch the parcel, does the light stay put. Those checks still matter. But a shot is only as useful as the cut it makes possible. If you control where a motion ends, and where the next one begins, you can plan that cut before you spend any credits.

What an editor needs from a shot

In chapter two, the difference between a shot that looks good and a shot an editor can use came up briefly. Now it does real work, so it needs plain terms.

A cut is the point where one clip stops and the next begins. The editor decides the exact frame. To make that choice, the editor needs spare material on both sides of the moment they want. That spare material is called a handle. The head handle is the footage before the part you plan to use. The tail handle is the footage after it. Handles let you nudge a cut a frame earlier or a frame later until the movement lines up. A clip with no handles forces you to cut at one exact spot, whether that spot works or not.

Overlap is related but different. It means the same action shows up in both clips. On a real film set, an actor repeats the movement for each camera angle. Mara would stand up in the wide shot, and then she would stand up again for the closer shot. Because the rise appears in both takes, the editor can leave the wide shot partway through the rise and pick it up in the closer shot at the same point in the movement.

That leads to the cut you are aiming for here, called a match on action. You cut while the motion is happening, not after it stops. Say Mara is halfway up from her crouch. You leave the first shot there and enter the second shot at that same halfway point. The viewer's eye follows the moving body across the cut and barely registers the change. It feels like one continuous rise seen from two places.

This is why the second shot has to begin a little before the first one ends. If the new clip started with Mara already standing, the rise would go missing. The editor would have nothing to match. If it starts while she is still crouched, the two clips share a stretch of movement, and the cut can land anywhere inside that stretch.

Alignment inside that stretch is not forgiving. Cut into the second clip too far along and a few frames of the rise vanish, so Mara seems to jump upward. Cut in too early and some of the movement plays twice, and she seems to stutter. Handles are what let you find the frame between those two mistakes.

The match on action has two more conditions. The camera has to change enough between the shots. A common working rule is at least thirty degrees of angle, or a clear change in framing size such as wide to medium. If the second shot sits almost where the first one did, the cut reads as a glitch, which is called a jump cut. The camera also has to stay on the same side of the action, so that if Mara moves toward the right of frame in one shot, she keeps moving toward the right in the next. Editors call this the hundred-and-eighty-degree rule. Picture a line running through the action. The cameras stay on one side of it.

Now put the accepted clip against these conditions. It is a wide, eye-level shot of the step, and it ends with Mara crouched and her hand just leaving the parcel. So the next shot should start in that crouch, begin the rise almost at once, and be framed differently. A lower, closer medium shot works well, with the camera on the same side of the step. It should end with Mara upright and turned toward the street, ready for whatever comes next.

That is a shot defined by both of its ends. Image to Video, from chapter three, lets you fix only the first frame. The ending is left to the model. For this job you need to fix the ending too.

First-and-last-frame control in Runway

First-and-last-frame control, also called keyframe generation, works like this. You give the model two stills: the first frame of the clip and the last frame. The model invents the motion that carries one into the other. Your prompt describes that motion. It does not describe the pictures.

In Runway, this lives in a separate application called Animate Frames. You find it under Runway's Apps and Workflows area, not in the main Gen-4.5 generation screen from the last two chapters. It needs a paid plan: Standard, Pro or Unlimited. A free account cannot run full multi-frame jobs. Runway's Workflows also include some models from other companies that offer keyframe modes, among them MiniMax's H3. For this project, Animate Frames is the native route, so the courier work stays with it.

The inputs are stills in PNG, JPEG or WEBP format, each up to roughly eight to ten megabytes. The output takes its aspect ratio from the frames you supply. Supported shapes include 16:9, 9:16, square, 4:3, 3:4 and the very wide 21:9. One rule matters more than the rest. Both frames must share the same aspect ratio. If they do not match, expect stretching or cropped edges. The courier project is widescreen, so both frames will be 16:9.

Native Animate Frames clips run five seconds or ten seconds. There is no three-second option like the one you used in Image to Video. That is fine here, because you want spare material, and the extra seconds become handles. Standard generation costs about fifty credits for five seconds and a hundred for ten. A faster Turbo pipeline costs about half that. The MiniMax H3 option allows five to fifteen seconds, billed per second, with a higher rate at its larger resolution. Every one of these jobs takes credits from your pool, even on an Unlimited plan. Prices on a live service change, so read the cost the interface shows before you press generate.

The output is silent. That matches the Image to Video clip, and the ambience file already in the project folder will cover both shots on the timeline. For comparison, Google's Veo 3.1 offers a similar First and Last Frame to Video mode. It makes four-, six- or eight-second clips and generates synchronized sound along with the picture. That is a real difference if you want audio from generation. It does not solve this shot's problem any better, though. The picture continuity is the hard part, and Runway handles that. So the project's tool stays the same.

Runway's advice on prompting follows from how the method works. The two frames already show who is in the shot, what they wear and where they stand. So do not describe any of that again. Spend the prompt on what happens between the frames: the physical action, the timing, the camera path and any change in the surroundings. Name camera moves in plain film terms, like "slow push-in" or "pan right." Keep the lighting, the scale of the subject and the horizon line consistent across both frames. When the two frames disagree on those things, the model tries to bridge the gap by warping the picture. That warping is where morphing and strange artifacts come from.

The demonstration: building the rise

What follows is an illustrative walk-through. The takes, the choices and the costs are examples that show the method. They do not describe anything you have done.

Step one is the start frame. There are two sources. The first is the accepted crouch clip itself. Its last usable frames show Mara's exact pose, her clothes and the light at that moment. The second is the 16:9 reference still of Mara from chapter three, which fixes her face and wardrobe more reliably than any single video frame can.

To pull one frame out of a clip, you use Shotcut's frame export. Open the project and move the playhead to the frame you want. The left and right arrow keys step one frame at a time, so you can land exactly. Then choose Export Frame from the File menu, or press Control, Shift and E on Windows or Linux, or Command, Shift and E on a Mac. A save window opens. Pick the project folder, give the file a clear name, and choose PNG so the image is not compressed again. One condition is easy to miss. The exported image includes whatever filters are active on the clip and the current preview scaling. If preview scaling is set to a lower resolution to keep playback smooth, you will export a smaller, softer frame. Turn preview scaling off before you export.

Which frame should you choose? Not the very last one. Pick a frame about half a second before the end, when Mara's hand has just left the parcel and her weight is starting to move. That is where the overlap begins. The second shot has to start there, a little before the first one ends.

That frame is from the wide angle, though. Using it as the start frame of the new shot would continue the old shot rather than create a new one, and the cut would be a jump. So the exported frame becomes a pose and light reference, not the start frame itself. Using the same image tool and the same reference still from chapter three, prepare a new 16:9 image. It shows Mara in that exact crouch, seen from lower and closer: framed from the knees up, parcel at the bottom of frame, camera on the same side of the step. The key light comes from the same side as in the wide shot, because the sun has not moved.

Check that frame before going on. Compare it with the exported wide frame, side by side. The collar should lie the same way. The bag strap should be on the same shoulder. The hand should sit in the same place relative to the parcel. Anything that differs here will show up as a continuity error at the cut, and that is far cheaper to fix in a still than in a clip.

Step two is the end frame. Mara stands at full height, turned toward the street, one foot beginning to step off. Make it from the same reference still, in the same medium framing and at the same camera height as the start frame, with the key light on the same side. The camera will move a little, so the framing need not be identical. But keep the horizon at the same height and keep Mara's size in frame close, because big jumps in scale between the two ends are what bring on morphing. Save both frames as 16:9 PNG files in the project folder, named so their order is obvious.

Step three is the prompt. It carries only the motion. A usable version: "She pushes up from the crouch, weight shifting onto her back foot, rising steadily, then turns her shoulders toward the street and begins a step. Slow tilt up following her rise." Look at what it leaves out. There is no description of her face, jacket, the bakery or the light. Both frames already show all of that. The words spend their effort on how the body gets from one frame to the other, and on the camera.

The phrase "rising steadily" is there on purpose. The model has five seconds to fill and only two poses to connect. Left alone, it may hold the crouch for most of the clip and then cram the whole rise into the last second. Naming the timing pushes the movement to spread across the clip.

Step four is the cost check. In Animate Frames, choose five seconds and the standard pipeline. The interface should show about fifty credits. Five seconds is enough: roughly the first half second overlaps the wide shot, the rise and turn take about two and a half seconds, and what is left becomes tail handle. Ten seconds would double the cost and leave the model more empty time to fill badly.

Step five is inspection. This method creates its own set of failures. Watch for four in particular.

The first is an implausible path. The model knows the start and the end but has to invent everything between them. Sometimes it picks a route no body would take, like standing up without the legs doing any work.

The second is a rushed middle. The action bunches up in one part of the clip, with a long still stretch and then a snap.

The third is a morph instead of movement. Rather than a person rotating, the picture dissolves from one pose into the other, like two photos blending together. Watch the face and the hands during the turn. Morphs hide there.

The fourth is lighting that does not match at the join. The first frame of the new clip has to sit against the wide shot. Even if both look fine separately, a shadow that falls a different way, or a warmer or cooler tone, will jump out at the cut.

Watch the clip at full speed first, the way a viewer would. Then step through the rise and the turn frame by frame with the arrow keys.

Step six is the rejected take. In the illustrative first take, Mara holds the crouch for nearly three seconds, then rises in under one. Her knees barely bend as she comes up, as if something lifts her. During the turn, her face goes soft for about four frames and her profile changes. That is a morph. The reason for rejecting it: rushed middle, a rise with no weight in it, and identity lost during the turn. Every one of these problems falls inside the stretch the cut needs, so no trim can save it. Write that down, along with the fifty credits it cost.

The fix is aimed at the causes. The rush suggests the model did not know when to begin, so the revised prompt opens with "She immediately pushes up from the crouch." The morph suggests the gap between the two poses was too big for a clean turn. The end frame had her shoulders turned almost fully away. Prepare a new end frame with her turned about three-quarters toward the street, so the face stays partly visible and the model has less rotation to invent.

Step seven is the accepted take. The illustrative second take starts the rise within the first few frames. Her back foot takes her weight, her knees extend in a believable arc and the tilt follows her up smoothly. Her face stays hers through the turn. At the start of the clip, the key light falls from the same side as in the wide shot, and the shadow under the parcel matches. It is accepted for those reasons: the rise has weight, identity holds and the light matches at the join. Two takes at fifty credits each makes a hundred credits for one accepted five-second clip. Only about three seconds of it will end up in the cut. So this shot costs about thirty-three credits per usable second. That is the honest number to write down.

Cutting the two clips together

Import the accepted clip into the Shotcut project. If Shotcut offers to convert it to an edit-friendly copy, accept, just as in chapter two. Generated clips often have a variable frame rate, and a constant-rate copy prevents dropped or duplicated frames, which would ruin a cut judged frame by frame.

Place the new clip on the timeline right after the crouch clip. Now find the moment to cut. In the wide shot, step forward frame by frame until Mara's weight first starts to lift. That frame is the earliest point the cut could happen. Then look a little further, to the frame where the rise is clearly under way, roughly a third of the way up. That is where the wide shot ends. Drag the tail of the wide clip back to that frame, or put the playhead there and press S to split, then delete what follows.

Next, trim the head of the new clip. Step through its opening frames and find the one where Mara's body is at the same height, with knees bent the same way and her hand at the same place. Drag the head of the clip forward to that frame. The half-second of overlap you built in is what makes this possible. Both clips contain the start of the rise, so you can choose the matching point from either side.

Now judge the cut, not the clips. Play from a second before the cut to a second after it, at full speed, several times. Does the rise feel like one continuous motion? If Mara seems to jump up, you have dropped frames, so move the new clip's head back by one or two. If she seems to repeat a sliver of the rise, move it forward. Adjust one frame at a time, because a single frame is often the whole difference. Check the light at the join and check that she keeps moving the same way across the screen. Then play the whole sequence from the parcel on the step to the turn toward the street, with the ambience running under both clips.

Save the Shotcut project, and save the records beside it: the exported pose frame, both start and end frames including the rejected end frame, both prompts, the settings (Animate Frames, five seconds, standard pipeline, 16:9), the credit costs of both takes, and the rejection note with its reasons. Those records are what let the shot be changed later without starting over.

This pair of clips is the first real step toward a multi-shot scene. Two angles hold one character through one continuous action, and the cut between them was planned before a single credit was spent.

A Runway plugin inside DaVinci Resolve Studio

One recent change bears directly on the kind of work this chapter did. On September twenty-third, twenty twenty-six, Runway released a plugin for DaVinci Resolve Studio. It brings Runway generation into Resolve's timeline, so you no longer export frames, upload them in a browser and download the results.

The inputs will look familiar now. You mark an in point and an out point on the timeline to set the range to work on. You supply reference images or start frames and a text prompt, and you can add up to five control keyframes. The plugin adds a workspace powered by Runway's Aleph two model. It can restyle footage from an anchor frame, keep the layout of a scene in place and hold keyframe continuity across clips on the timeline. In practice, much of the round trip you just did by hand could happen in one place: pull a frame, prepare the ends, generate, return to the cut.

The conditions matter. It needs DaVinci Resolve Studio, the paid edition, version nineteen or later, on a Mac or Windows computer with an internet connection. The free edition of Resolve is not supported. The installer costs nothing, but every render takes credits from a paid Runway plan at Standard or above. So it adds to your costs rather than replacing them.

For this project, it does not mean switching editors today. Shotcut still does the job, and chapter two explained why Resolve's heavier hardware demands and layout wait for later. The smallest useful next action is simple. If you already own Resolve Studio, check that it is version nineteen or later, and note the plugin as a candidate for the finishing stage. Either way, keep saving start frames, end frames and prompts as separate files, as this chapter did. Frames and prompts stored like that move easily into any tool that accepts them, including this one, once a real finishing need gives it a job.