Skip to main content
The hard part of an AI cutscene is not making a good shot. It is making three shots in a row where the hero has the same face, the same cloak and the same light, so the player believes it is one moment. Ask a video model for each shot from words alone and you get three different people. Ask for one long take and the face drifts after six seconds. The fix is the method film crews use: lock the look in still pictures first, then animate each picture as its own short shot, then cut them together with a voice line and a music cue. This guide makes a 15-second story beat that way, from nothing to a finished MP4.

​
The chain

$2.721 for the whole chain, at each model’s list price:
  • Character reference picture: $0.053 ($0.053 per image)
  • Style frame (shot 1): $0.069 ($0.053 per image)
  • Start frames, later shots x 2: $0.08 ($0.04 per image, 2 jobs)
  • Image-to-video shots x 3: $2.355 ($0.157 per second, 3 jobs)
  • Sound for each shot x 3: $0.105 ($0.035 per clip, 3 jobs)
  • Voice line: $0.0103 ($0.08 per 1,000 characters)
  • Music cue: $0.0425 ($0.0025 per second)
  • Sound accents x 2: $0.006 ($0.002 per second, 2 jobs)
A failed job costs nothing. Before you run anything, price each step exactly and for free: summer api estimate <model> --input '<json>', or POST /v1/estimate with the same body as the job. Prices come from GET /v1/models and change only when the provider’s price does.

​
Before you start

Sign in once, so your agent or your terminal can run jobs on your Summer account:
Your browser opens the Summer sign-in. Pick a spend limit for this computer and choose Allow. Using a coding agent? Paste this guide’s URL and say “make this cutscene for my game”; the quickstart for your agent connects it in one command. Then write the shot list. A cutscene longer than about 6 seconds is several shots, never one take. One line per shot: framing, the one action, the one camera move. Three 5-second shots make a 15-second beat. A still costs a small part of what a shot costs, so do all the fixing on stills.

​
Step 1: a character reference picture

This picture is the character’s identity for the whole cutscene. Every later frame is an edit of it or of a picture made from it. Make it like a casting photo: the full figure, front view, plain background, even light, and every detail named in words (“moss-green hooded wool cloak”, “brass oil lantern”). Words carry colours and materials into the edits better than pixels alone.
If the character is already in your game, start from a screenshot or the concept art instead: upload the file and use its fileUrl in step 2. The 3D character guide makes a concept picture that works here as it is.

​
Step 2: the style frame

The style frame is shot 1’s first frame, and it sets the look of the whole scene: the light, the palette, the lens. Edit it from the character reference with referenceImageUrl, and say “same face, hair, cloak and lantern as the reference”.
Make it 16:9, or 9:16 for a phone game. An image-to-video shot keeps the framing of its start frame, so a square start frame gives you a square shot whatever aspectRatio says. Look at the style frame for a full minute before you go on. A wrong colour here is wrong in every shot.

​
Step 3: start frames for the other shots

Make each later shot’s first frame as an edit of the style frame, not of the character reference. The style frame already holds the character and the light, so both carry into the close-up and the over-the-shoulder shot.
Run the step once per shot, with that shot’s framing and action from your list. The second frame, for shot 3, is: “Over-the-shoulder shot from behind the same woman, Wren, her raised lantern in the foreground, looking down the foggy street where a tall shadow stands between the dark lanterns.” Keep the closing sentence (“same face, hair, cloak, lantern, light and colours as this picture”) in every frame prompt, word for word.

​
Step 4: animate each shot

Now each start frame becomes a 5-second shot. The prompt describes only what moves: one action and one camera move. The picture already says what everything looks like, so do not describe the character again; repeated description makes the model redraw her.
Run it three times, once per start frame, and change the prompt to that shot’s action. What works:
  • One verb for the camera. Slow push-in, static, slow pan left. Complex moves through a crowd fall apart.
  • 5 seconds. Faces, hands and physics drift with every second past about six. Two 5-second shots look better than one 10-second shot and cost the same.
  • Describe only the picture. The next step makes the sound from the finished shot, so the prompt spends every word on the action, the light and the camera.
  • A negative prompt for the usual faults. changing face, extra fingers, warped hands, text, subtitles.
A video job takes a few minutes. The CLI follows it and saves the MP4; over HTTP you read GET /v1/jobs/{id} until state is done (jobs). Try a new shot first on a cheap model such as MiniMax H3 Max Turbo ($0.026 per second) to check the motion, then make the final. Other models for this step. Kling 3.0 Pro makes sound with the picture in one job ($0.168 per second): ask for it in the prompt (ambient sound only: soft wind, a lantern creaking, then no speech, no music) and skip step 5. Veo 3.1 gives the most coherent shots, also with sound, for the one shot that must be perfect ($0.40 per second). HappyHorse 1.1 makes the character speak with lip sync when you write the line into the video prompt: the voice then comes from the video model, not from your voice cast. To keep a character the same across many shots without a start frame, the reference models take up to several pictures of her: Kling 4.0 References ($0.157 per second).

​
Step 5: sound for each shot

Send each finished shot back as videoUrl and describe what should be heard. The model watches the clip and makes sound that follows it: a step lands when her foot does, the flame crackles when it flares. Ask for the place and its things, and keep No speech, no music: the voice and the music come from their own jobs.
The answer is the same clip with its new soundtrack. ThinkSound ($0.01 per second) reasons about the scene before it makes the sound, and suits slow, quiet shots where every small sound counts; it takes longer. Kling Video to Audio ($0.035 per clip) is fast and fits most shots.

​
Step 6: the voice line

The line is its own job, so you choose the voice, the delivery and the timing. Pick the voice from the model’s list (npx -y summer-engine@latest api models voice/expressive-v4 prints every voiceId). Alice, a clear and calm voice, fits a lantern keeper who has seen something.
Stage words in square brackets set the delivery: [quietly], [pause], [whispering]. Full stops make clear pauses, and an ellipsis trails off. A video model does not move lips to a separate voice file. So play the line over shots where the mouth is small or hidden: the wide shot and the over-the-shoulder shot here. The voice lines guide covers casting and delivery in depth.

​
Step 7: the music cue

Time the cue to the scene: 15 seconds of picture plus a 2-second ring-out, so the last chord fades under the cut back to gameplay. At 60 BPM one bar is 4 seconds, so events on bar lines (0:08, 0:12) land where you ask more often. The swell starts under shot 2 and peaks inside shot 3.
Make two or three takes and keep the one whose peak lands nearest 0:12. Then move the cut a few frames to the music, not the music to the cut. For a score that shares one theme across the menu, the cutscenes and combat, see music for your game’s cutscenes.

​
Step 8: sound accents on the cut

The shot sound gives you the air of each shot. It does not put a sound on an exact frame. For the one or two moments that must hit (the lantern flaring on the cut to shot 3, a distant bell), make them as sound effects and place them yourself.
For a long scene, add one loop: true ambience per location under the shots, at the length of the scene, so the air never drops out between cuts.

​
Assemble the shots

Cut on action: end shot 1 as she starts to turn, open shot 2 mid-turn. Then lay the sound in four layers: shot sound low, the voice on top, the music under it, the accents on their frames. In any video editor that is four tracks. From a shell, two commands do it:
The voice starts 2 seconds in (adelay=2000), the shots’ own sound sits at 35 percent and the music at half volume under the voice. The concat step needs every shot to have a sound track; with a silent video model, drop [0:a] and mix your ambience loop in its place. Add the accents the same way, each with its own adelay.

​
Put it in your game

  • Summer Engine. Import the MP4 by its assetId or URL; the import converts it to .ogv, the video format the engine plays, and names the new file. Play it in a full-screen VideoStreamPlayer with a short fade from and to black. The cinematic cutscene skill builds the scene, the fade and the hand-back to gameplay for you.
  • Any other engine. The MP4 plays in any video player node. Pause gameplay, fade in, play, fade out, and let the player skip.
  • Keep the stills. The character reference and the style frame are now the canon for this character. The next cutscene edits from them, so Wren looks the same in chapter five.

​
Make every cutscene in the game

Once the first scene works, the next one is a batch. Write every start frame for the scene into one JSONL file and send it with summer api run --batch frames.jsonl --download cutscenes/ch2. Review the stills, then send the shots as a second batch, each with its frame’s fileUrl as imageUrl. A batch is priced up front and runs all or none (batches).

​
Common problems

​
Questions

Can I make a cutscene from my own game’s art? Yes. Upload a screenshot or the concept art (send a picture) and start at step 2 with its fileUrl. How long can one shot be? duration goes up to 20 seconds, and each model renders its nearest supported length. For a character on screen, 5 seconds per shot holds the look best. Can I add sound to a video I already have? Yes. Upload the clip (at most 20 seconds) and run step 5 on it with its fileUrl as videoUrl. That works for gameplay captures too. Can I use the cutscene in a commercial game? Yes. Every model on the Summer API allows commercial use of what it makes.