From Idea to Image, Part Seven

Prepared with Claude; edited and published by Owrang.

About this series

From Idea to Image is a weekly self-study course in making images with AI. The text of the series is also written with the help of AI and is edited and published by Owrang.

Each part teaches one skill and ends with a simple exercise you can do the same day.

Over the last six parts we built Mahin, a woman in her sixties with a green cardigan, round glasses and a white streak in her hair, who sits at the kitchen table holding a glass of tea. Last week we brought her to life. She breathed, blinked and turned her head towards the window. But each of those clips was a single shot. This week we put several shots side by side to build a small scene. A wide shot of the kitchen, a close-up of the glass of tea, Mahin's look towards the window, and what she sees outside it.

A good shot is not a scene

In ordinary filmmaking, a large team works together so that the shots of a scene belong to one world. The production designer builds the place, the costume designer and make-up artist build the character's look, the director of photography the light, and the sound team the world of sound. The script supervisor notes the details of every shot, from where the cup stands on the table to the hand the character used to open the door, and makes sure the shots match. We call this matching continuity.

When you work with AI, all of these jobs are yours. And the tool does not remember the previous shot. Each time you make a new shot, the tool imagines the scene from scratch. It may move the window to the other side, send the light from another direction, change the colour of the cardigan slightly or put more tea in the glass. None of this shows in a single image. But when you watch two shots one after the other, the scene falls apart.

So this is the lesson for the week. Design the scene before you make the shots, and judge every new shot against the others, not on its own.

Before the map, your own vision

Tools today offer plenty of ready-made modes and templates, from sample prompts to modes that plan the shots by themselves. A template makes the work faster, but it does not necessarily make creative work. What turns a few shots into a work is the vision of the person making it. The AI carries it out, and you lead.

So before you draw the map, write one sentence. What is this scene about? Why does Mahin look at the window? Perhaps she is waiting for someone. Perhaps she is watching the clouds to see whether it will rain. Perhaps something has come back to her. Each answer makes a different scene. If she is waiting for someone, the shot of the door may matter more than the shot of the glass. If something has come back to her, perhaps the tea in the glass has gone cold and no steam rises from it any more. The shots we suggest in this lesson are one example, not a template to copy. Build your map, your shot list and your prompts from your own sentence.

Before the computer, with paper, pencil and phone

This week, before you write any prompt, do three small things.

First, a map of the kitchen from above. On a sheet of paper, draw Mahin's kitchen from above, as it would look if we had lifted off the ceiling. Draw the table as a rectangle, Mahin as a circle, the window as a line along the wall, and the door wherever you like. Now draw a straight line from Mahin towards the window and extend it to both edges of the paper. This is the imaginary line. Put all the cameras of this scene on one side of this line, and mark each one with a small triangle pointing at what it sees. If the cameras stay on one side of the line, the window stays on the same side of the frame in every shot and the viewer does not get lost. This is called the 180-degree rule.

Second, four frames. Draw four small rectangles side by side and in each one show, with a few simple lines, what is seen. The wide shot of the kitchen, the glass of tea up close, Mahin's face as she looks at the window, and the window. Under each frame write how many seconds it lasts. This sheet is your shot list, and from now on you will check every clip against it. In cinema this is called découpage, breaking a scene into shots before filming.

Third, the sound of the room. With your phone, record fifteen seconds of sound in your own kitchen while nobody is talking. The hum of the fridge, the ticking of a clock, a car passing in the street. You will put this sound under every shot of the scene. It is this continuous sound that turns four pieces of picture into one moment. Absolute digital silence between shots sounds like a hole in the scene.

If you would like one more exercise, film yourself on your phone picking up a glass, once from a distance and once up close, and each time repeat the whole movement from start to finish. Then, in your phone's editor, cut the distant shot right in the middle of the hand's movement and continue with the close shot from the same moment of the movement. You will see that a cut in the middle of a movement is almost invisible. This is called cutting on action.

Today's technology, from shot to scene

In the last few months video tools have gained new features that help with building scenes. What follows comes from the makers' own official pages. But tools change every month, and sometimes they disappear. According to OpenAI's help page, the Sora app and web version were closed on 26 April 2026 and its API on 24 September 2026. That is why this series teaches principles, not the buttons of one particular tool.

First frame and last frame. Last week we gave the tool only the first frame. Now more tools also accept a last frame and make the clip between the two images. Google added this to Gemini Omni Flash 1.1 on 27 August 2026, and Kling 4.0, available to everyone since 15 October 2026, lists it among its features. This is very useful for building scenes. If the last frame of one shot is the same pose the next shot starts from, the two shots meet at the cut.

Several shots in one request. Some tools now make several shots, with cuts between them, in a single generation. According to the Kling 3.0 guide, in Multi-Shot mode the model plans the cuts itself, and in the custom mode you can set the number and length of the shots, in a clip of 3 to 15 seconds. The Gemini Omni Flash guide also says this model makes a video with a few different shots by default, and that if you want one continuous shot you must say so explicitly, for example "in a single continuous shot, no cuts". The feature is tempting, but with it the tool decides about the cuts. For this week's exercise, make the shots one by one so that the découpage stays in your hands.

References that keep a character the same. New tools accept several reference images of a character to keep it the same across shots. The Kling 3.0 guide asks for two to four reference images to build a fixed character or object, which it calls an element. Google, in its best-practice guide for video generation, recommends copying the complete, unchanged description of the character into the prompt for every new scene. This is what we do in this series with Mahin's description and her character sheet.

Extending a clip. According to Google, Omni Flash 1.1 can extend a video in ten-second steps up to forty seconds. Extending makes a longer shot, but it does not make a scene. A scene is made by putting shots side by side and by where you cut.

Sound and dialogue. Several tools now generate sound and dialogue together with the picture, and Kling 4.0 reports more accurate lip sync with speech. But Persian does not appear in the list of languages on any of the official pages we checked. The Kling 3.0 guide says that if dialogue is written in a language other than its five supported languages, the model translates it into English. The Gemini Omni Flash guide says English is fully supported and other languages have not yet been evaluated. So for a scene in Persian, record the sound separately and lay it under the picture in your editor. Next week we will turn to exactly this.

One more point, about your own contribution. The US Copyright Office wrote in its report of 29 January 2025 that, with current technology, prompts alone do not give enough control for authorship, but that the creative selection, coordination and arrangement of material, and creative changes to the output, are the human's contribution. Editing a scene, choosing the shots, their order and where to cut, is exactly this kind of work.

This week, step by step

Step 1. The character sheet. Before you make any new shot, give the image tool Mahin's reference image and make a character sheet of her, three images of the same woman in the same clothes and the same light, side by side on a plain background. Front, three-quarter and profile. Take the base of your prompt from the earlier parts and add this to it. "Three views side by side of this same woman on a light grey background, front, three-quarter and profile, standing in even light, with the same green cardigan, round glasses and white streak in her hair." From now on, give this sheet to the tool with every prompt. The more angles from which the tool has seen Mahin, the less it guesses in a new shot.

Step 2. The wide shot. This shot tells the viewer where everything is. Write it from your map. "Wide shot of the kitchen at eye level. Mahin sits at the wooden table on the left of the frame. The window with a thin curtain is on the right of the frame, and the soft light of an overcast morning comes through it. The kitchen door is behind her, in the depth of the picture. The camera is fixed on a tripod." Always state left and right explicitly. The tool does not remember them from the previous shot. In Mahin's reference image the light comes from the right, so we have put the window on the right. If in your images the window is on the other side, write that side and keep it in every shot.

Step 3. The insert, the glass of tea. A close shot of the glass on the table. "Close-up of a small glass of tea on the wooden table. A thin wisp of steam rises slowly. Soft light comes from the right of the frame." Keep the hands out of the frame. Last week you saw why. Watch the level of tea in the glass. If the glass is half full in the wide shot, it must be half full here too. This is exactly what a script supervisor notes down.

Step 4. The look. You do not need to make this shot from scratch. The clip from step 5 of the last part, the turn of the head towards the window, is this shot. If that clip came out well, use it. Just check that in that clip Mahin looks towards the right of the frame, the side where the window is in the wide shot. If she looks to the left, this shot does not fit your map.

Step 5. The window. What Mahin sees. "Shot of the kitchen window from beside the table, from the point of view of a seated woman. The thin curtain moves gently in a light breeze. Outside, an overcast morning sky and the branches of a leafless tree. The camera is fixed." This is called a point-of-view shot. Having seen Mahin look, the viewer sees what she sees, and the two shots together make one meaning.

Step 6. Assembling the scene. Put the clips in a simple editor. Your phone's own editor or any free program you know is enough. A suggested order is this. The wide shot, three seconds. The glass, two seconds. Mahin's look, with the cut right in the middle of her head turn. The window, three seconds. And back to Mahin's face for a moment, calm. If you like, swap the shots around and see how the meaning changes. The same calm face says something after the shot of the window that it does not say after the shot of the glass.

Step 7. Sound. Lay the room sound you recorded on your phone continuously under all the shots, from the beginning of the scene to the end, without breaking it at the cuts. If the clips have their own sound, lower it or remove it. If you want to add the sound of the steam or the curtain, let it be heard a little before the cut to its shot. When sound runs across a cut, before or after its picture, this is called a J-cut or an L-cut.

Step 8. The continuity log. Write one line for each shot. Which side of the frame is Mahin on? Which side does the light come from? Glasses, white streak, colour of the cardigan, gold chain, level of the tea, steam. Watch the scene from start to finish and mark whatever has moved.

When the scene does not hold together

Mahin is on the left in one shot and on the right in the next? State left and right explicitly in the prompt, and give the tool the wide shot as a reference image as well. Resist the temptation to mirror the image. Mirroring also moves the white streak to the other side of her head and turns everything the opposite way.

The light comes from the left in one shot and from the right in another? Write the direction of the light in every prompt with the same words. "Soft light from the window on the right of the frame."

Something appeared in the frame that nobody asked for? Watch every clip to the last frame and to the edges of the frame, with sound, before you let it into the scene.

Writing appeared on the window glass, on a picture on the wall or on a book cover? Tools do not write letters correctly and produce meaningless text. Make the shot without writing.

The sound of the shots does not match? Remove the clips' own sound and put one continuous room sound under all of them.

You made a shot several times and it did not work? Do not make it a third time. Go back to the map and the shot list. Perhaps the shot needs to be smaller, perhaps the angle needs to change, perhaps the job can be done with a cut instead of a movement.

What happened in practice

Owrang works on visual projects with AI in which a character has to be seen and heard, across several shots, like a real human being. What we learned in this work shows this week's lessons better than any made-up example.

First, new shots did not match old ones. In a project part of which had already been edited, we tried to replace a few shots with newly generated ones. Each new shot was good on its own, but next to the old shots the light, the colour and the details of the costume did not match. In the end we decided to work on the original shots themselves instead of making them again. The lesson was to make all the shots of a scene from the start with the same references. Matching them later is far harder.

Second, someone nobody asked for. We asked a tool only to turn a character's head slowly towards the crowd. The head movement came out natural, but in the corner of the frame the tool created a woman who stood facing him and talked, with a voice nobody had asked for. Since then we watch every clip to the last frame, to the edges of the frame and with sound before it goes into the edit.

Third, one word in the prompt. The prompt for one shot said that the character was speaking and asking the crowd a question. His mouth opened and closed for no reason, and a man who should have been calm looked like someone swearing at everyone. The words of a prompt direct the movement, like a director talking to an actor. From then on we read the prompt for every shot before making it and wrote only what we wanted. For a silent reaction we wrote "lips closed".

Fourth, lips have to be seen in motion. The still frames of a dialogue shot looked right, but in motion the jaw, teeth and lips lost their place. Once the lips were also a frame or two apart from the sound. The way to fix it was to move the sound by one or two frames in the editor. At 24 frames per second, each frame is about 42 thousandths of a second. Judge a dialogue shot only by watching and listening at the same time.

Fifth, a silence that has sound. We made silent reaction shots for the character, before and after he spoke. The working rule became never to put absolute silence under these shots. The sound of the scene's own space, the crowd and the wind, has to continue under every shot.

Sixth, a shot made in reverse. We wanted a reaction shot to end exactly in the pose the next shot starts from. The tool starts a clip from the frame you give it, but its end cannot be set precisely. So we made the clip starting from that pose and played it backwards, so that it arrived exactly at that pose. This trick only works for movements that also look natural in reverse, such as turning the head. The eye notices steam, smoke and walking in reverse at once. At the join, too, the generated clip did not start exactly on the given frame, and a short dissolve, under a quarter of a second, covered the seam. Today, tools that also accept a last frame have made this easier.

Seventh, signs without sources. In historical scenes the tool made patterns, writing and ornaments that had no source behind them and that looked completely believable in the natural light of the image. Our rule became that everything in the frame must either have a source or a job in the story, and that wherever no source was found, we say plainly that none was found.

Working terms

Découpage, breaking a scene into shots before making them, découpage / shot breakdown

Shot list

Storyboard, drawn frames of the shots in order

Establishing or wide shot, a shot that shows where everything is

Insert, a close shot of an object or detail

Point-of-view shot, what the character sees, POV

Reaction shot

The imaginary line, the 180-degree rule

Eyeline match, the direction of a character's look matches what is seen in the next shot

Cut on action

Continuity

Room tone, the sound of a space without dialogue, ambience

J-cut and L-cut, sound that runs across a cut before or after its picture

Character sheet

Dissolve, one shot gradually fading into another, cross-dissolve

This week's exercise

Build a small four-shot scene for your own character. Before anything else, write in one sentence what this scene is about. Then draw a map of the place from above, mark the 180-degree line, put the four frames of your shot list on paper and record fifteen seconds of room sound. Then make the character sheet and from it make a wide shot, an insert, a shot of the look and a point-of-view shot. Assemble the scene in an editor, put the cut of the look on the action and lay the room sound continuously under everything. At the end, fill in the continuity log.

Check it yourself

Show the scene once without sound and once with sound to someone who has not followed the exercises. Ask them which side of the character the window was on, and whether all the shots happened in one moment and one place. Ask them what they think the scene is about, and compare their answer with the sentence you wrote before you started. Ask which version felt more continuous, with sound or without. If they placed the window correctly and found the version with sound more continuous, you have made a scene, not four shots.

Next week. We plan to turn to sound and dialogue. Mahin will say a short sentence for the first time, and we will see how to bring sound, lips and look together.

Educational links

All these links go to the official channel of the educational institution or publisher itself. Watch and learn, but do not republish anything from them and do not give them to a video tool. All these videos are in English and have automatic subtitles.

Shot sizes. Khan Academy with Pixar, from the Pixar in a Box series, "Basic shot types", wide, medium and close shots, and when to cut in closer to give a moment weight, about four and a half minutes. https://www.khanacademy.org/computing/pixar/storytelling/film-grammar/v/basic-shot-type

Shooting for the edit. BBC Fresh, the BBC's channel for new filmmakers, "Shooting sequences for the edit", a small scene in wide, medium and close shots and inserts, with the action repeated for cutting on action, nearly four minutes. https://www.youtube.com/watch?v=gJm5wtscceg

The imaginary line. BBC Fresh, "Crossing the line", the 180-degree rule and direction of movement in the frame, under two minutes. https://www.youtube.com/watch?v=Q-ze5c_Pz6s

Storyboarding. Khan Academy with Pixar, "Storyboarding", a Pixar story artist breaks down a scene from Ratatouille shot by shot, over-the-shoulder shots, reverse shots and cutting on action, about twelve minutes. https://www.khanacademy.org/computing/pixar/storytelling/film-grammar/v/storyboarding-scene

Continuity editing. The Open University, OpenLearn channel, "'Continuity editing' – Digital Film School (12/18)", an editor cuts six minutes of making tea down to thirty seconds and matches the movement from one shot to the next, about fourteen minutes. The video software is old, but the principles are the same. https://www.youtube.com/watch?v=G04_79CJqv4

The Kuleshov effect. CrashCourse with PBS Digital Studios, "Soviet Montage: Crash Course Film History #8", the experiment in which the same still face reads as hunger, grief or longing depending on the shot beside it, about twelve and a half minutes. https://www.youtube.com/watch?v=-RtBAa4YCgo

The sound bridge. Criterion, "Observations on Film Art: Sound in Fritz Lang's M", sound that carries from one scene into the next, off-screen sound and silence, about two minutes. https://www.youtube.com/watch?v=My7fxjUY2co

The editor's work. CrashCourse, "The Editor: Crash Course Film Production with Lily Gladstone #12", from the first assembly to the final cut, and how choosing the moment of a cut makes a scene belong to one character or another, about eleven minutes. https://www.youtube.com/watch?v=esa0NeQI8oc