A short film with AI: the pipeline from script to final cut

What goes wrong

Before you start

• A script with named beats. Five beats is twenty five seconds.
• One sentence per character, per place and per prop, written the way a costume or art department would write it. These sentences are pasted, unchanged, into every prompt.
• A NOLGIA account. Characters, Locations, the composers and Studio come with it.
• Patience for one redo per stage. Ours needed five, and the guide says where.

LAST TRAIN. Night. An underground platform, magenta neon over white tiles. 1. Vale runs for the last train, a chrome case in one hand. 2. She passes Ozzy, the busker, who turns to watch. 3. The doors close in her face and the train pulls away. 4. The case, set down on the wet platform. 5. On the bench, out of breath, she laughs.

Save the cast, the set and the prop

Open Characters. Type a name and the character's one sentence as the description, open Add reference images, and either upload a photo or use Generate with a portrait prompt and a model. With the portrait selected as the reference, press Create character. We made Vale this way on GPT Image 2.5 Flare, and Ozzy from a still made the same way.

Full-length portrait of a woman in her late twenties, silver-blonde chin-length bob with a straight fringe, pale skin, red lipstick, wearing a glossy red vinyl trench coat belted closed over a black ribbed turtleneck, black wide-leg trousers and chunky white boots, holding a folded transparent bubble umbrella, standing on an underground subway platform with white tiles and a magenta neon strip along the ceiling, wet floor reflections, thin haze. Photoreal, cinematic, editorial fashion lighting, no text, no logos.

Open Locations and do the same for the place: a name, a description, the one sentence as the canonical description, and a still from your Library or an upload. Our platform is a Seedream 5.0 Pro still. The prop is a still too, made in the image composer and kept in the Library; it rides into an insert as a reference image and into every other prompt as one sentence.

An empty underground subway platform at night: glossy white tiles with one teal stripe at waist height, a long magenta neon strip along the ceiling, wet floor with mirror reflections, thin haze, a yellow safety line along the platform edge, a single steel bench, the tunnel mouth glowing cyan. No readable text or signs.

Make each shot as a frame first

Open the image composer. Pick GPT Image 2.5 Flare, set the shape to 16:9, then press Characters and pick Vale, and Location and pick Halcyon Street station. The composer shows a chip for each. Write the shot as one moment, name the character with an @ in the prompt, and press Generate images. Five frames, one per shot, cost less than one video take between them and are where the wardrobe, the props and the framing get fixed.

@Vale mid-sprint along the empty platform towards the camera, her red vinyl trench coat flying open over the black turtleneck, the folded transparent umbrella in her left hand and a small brushed-chrome hard-shell case with a black rubber handle and two black latches in her right, white boots splashing the wet tiles. Wide low shot, the magenta neon strip above her, thin haze, the tunnel glowing cyan behind. Cinematic film still, anamorphic, fine grain, no text.

Animate the frames into shots

Open the video composer, choose Animate an image under Start from, press Add start frame, then Pick an asset and choose the frame from your Library. It becomes the first frame of the clip. Write the action that happens from that frame, press a camera move chip, set the length, and generate. We used Kling v3 for the sprint, the shot with the most movement, and Seedance 1.5 Pro at 720p for the other four, 4 to 6 seconds a shot, and cut two of them shorter in Studio. For a run, a jump or anything with weight, animate the frame on Kling v3 or Seedance 2.5.

She sprints straight towards the camera along the platform, red vinyl coat flying open, the chrome case in her right hand and the folded umbrella in her left, white boots splashing the wet tiles, hair bouncing with each stride, the magenta neon and the cyan tunnel glow behind her. Real sprint physics, cinematic, anamorphic, fine film grain.

Score and effects

The film has no dialogue, so sound is two runs in the audio composer: a score from Lyria 3.5 and two effects from ElevenLabs Sound Effects v2, the doors closing and the running boots. Describe the score by tempo and mood, and each effect as one event.

A tense, driving electronic cue at 118 bpm for a short film about missing the last train: pulsing arpeggiated synth, deep sub bass, a ticking hi-hat, a sudden drop to a single warm pad at the end, no vocals, cinematic neon night

Subway train doors closing: a two-tone chime, a pneumatic hiss, doors thud shut, then the train accelerates away through a tunnel with a rising electric whine and echo, station reverb

Cut, grade and export in Studio

Open Timeline, press New timeline, name the project and the timeline, then open Library in the Media column and add the five shots in order with their + buttons: each press adds the clip at the playhead on a fresh track, and End moves the playhead to the end of what is there. For shot 2, drag the Source Preview to 5 seconds and press Mark Out before the +, so only its first 5 seconds come in. Add the score at the start; it is 2 minutes long, so park the playhead at the end of the picture, press C with nothing selected to split it, click the tail and press Delete. Add the two effects where the boots land and where the doors close. Then open Color from the toolbar with nothing selected and pick one film stock for the whole timeline; ours is CineStill 800T.

↔

BeforeAfter

Export video renders the MP4 in a minute or two; it lands in your Library at no charge. The arrow beside it opens the Export menu; Premiere Pro or DaVinci Resolve downloads one ZIP with sequence.xml, a media folder with every clip and track, your grade as luts/canvas.cube and a README. Unzip it, then import sequence.xml. Ours came with one warning per video clip: the XML assumes each clip matches the canvas size, so use Scale to Frame Size in Premiere if a clip looks wrong. The export post covers the rest of what carries over.

Which tool for which job

Where you still regenerate

• A model that refuses your cast. Seedance 2.0 Fast refused Vale's reference on every try. Switch models rather than rewording; the refusal names the checker, not your prompt.
• A portrait as a start frame. On Seedance 1.5 Pro a character attached directly becomes the first frame. Go through a 16:9 frame.
• The set repeating. The second take of shot 2 ran past the busker twice by 4 seconds. Shorter shots, or cut where it repeats.
• A prop that migrates or arrives. The first take of shot 2 gave Vale the saxophone by 2 seconds; the first take of the insert had the case fade in from nothing. Say who holds what in the shot prompt, and for a static prop put it in the frame you animate from.
• Faces. Wardrobe, hair and props hold across the five shots; whether the face is one person is not measured here.

What it costs

Frames are priced per image, shots by model, length and quality, the score and each effect per run. Studio, the cut and the export are free. Live rates for the models in this film:

Checklist

• Character, location and prop saved before the first frame; one sentence each, reused verbatim.
• One 16:9 frame per shot with the cast and the location attached; wardrobe and props checked at full size.
• Each frame animated at 4 to 6 seconds with one action and one camera move.
• Score by tempo and mood; effects one event each.
• Trimmed where the model drifts; one grade on the timeline; exported once.

Try it yourself

Save your own character and location, then use the frame 1 prompt with the cast and location attached on GPT Image 2.5 Flare at 16:9, and the shot 1 prompt on Kling v3 in Animate an image with the Locked off chip. We ran the frame 1 recipe again on the live composer after the guide was written: it came back on the same platform with the same coat, bob, boots, umbrella and case, in a different stride and with the coat belted closed, although the prompt says it flies open. Expect the pose and the small things to roll; the character and the place are what the references hold. The stills below are our cast, set and prop; the frames and shots are above.