How to make AI movies: block your scenes in 3D, with data from 9 films made in Flick

Yes, one person can make an AI movie: write the beats, pick one look, make reference sheets, plan each shot, turn stills into video, then cut with sound. Start each shot with a prompt, and when the prompt can't get the camera or the positions right, block the shot in Flick's 3D Stage: the agent builds the set from one of your images, colored shapes stand in for the actors, and you can film the camera move with your phone, so no 3D experience is needed. The 9 films made in Flick below took a median of 84 generations per finished minute.

The Only 3D-to-AI Filmmaking Tutorial You Need, with Jaime Yao (Flick on YouTube). Step 4 below is built on its method.

Films by Jaime Yao (Flick team and Flick Residency), Zoey Z. (Flick team) and Simay Ezer Pekcan (Flick Residency). Featured tutorial: Jaime Yao.

What it takes to make an AI movie: 9 films in numbers

Flick publishes a breakdown of each film's shared workflow: every image and video generated for it, including the takes that were thrown away. These are the 9 films with a public breakdown.

Film	Runtime	Images	Videos	Video generated	Top image and video model
In ’twixt Vows	16:00	1,145	373	90 min	Nano Banana Pro, Seedance 2.5
A Proper Introduction	5:45	897	217	32 min	Nano Banana Pro, Kling O3 Pro
The Herder	4:27	364	254	32 min	Nano Banana Pro, Kling 3.0 Pro
Ulysses	4:36	349	37	3 min	Nano Banana Pro, Kling O3 Pro
Embedding Valley	4:12	242	109	20 min	GPT Image 2, Seedance 2.0
In Our Time	3:47	33	98	9 min	Nano Banana, Seedance Pro
Sonny's worth	2:50	81	56	16 min	Nano Banana Pro, Seedance 2.5
To My Grandma	1:38	22	13	3 min	Midjourney, Seedance 2.5
The Little Fox’s Dream	1:31	59	59	9 min	Nano Banana, Seedance Pro

Flick film breakdowns, October 19, 2026. Counts include discarded takes; uploads and audio are not counted. Runtime from Flick's video sitemap.

• About 84 generations per finished minute. That is the median of the 9 films. It runs from 21 for To My Grandma, a narrated film built on a few simple stills, to 194 for A Proper Introduction, where every shot was sketched first and then edited many times.
• About 5.5 seconds of video for every second on screen. Across 8 of the films, the median film needed 5.5 times as much video as it runs, so most generated video never makes the cut. Fewer rerolls is what blocking a shot in 3D (Step 4) is for. Ulysses is left out here: its workflow holds 3 minutes of video for a 4:36 film, so its ratio doesn't compare.
• Most of the work is stills. 72% of all generations (3,192 of 4,408) were images: character sheets, frames and edits made before anything moved. A Seedance 2.0 or 2.5 take with references costs 10 to 60 times as much as an image (see the cost table below), so the films settle each frame first.
• In ’twixt Vows by Jaime Yao (Flick Residency)
• A Proper Introduction by Jaime Yao (Flick team)
• The Herder by Jaime Yao (Flick team)
• Ulysses by Jaime Yao (Flick team)
• Embedding Valley by Jaime Yao (Flick team)
• In Our Time by Jaime Yao (Flick team)
• Sonny's worth by Zoey Z. (Flick team)
• To My Grandma by Zoey Z. (Flick team)
• The Little Fox’s Dream by Jaime Yao (Flick team)

Step 1: Write the idea as beats before you generate anything

A beat is one moment the audience has to see, in order, with a rough length. Writing the beats first tells you which shots and references you need, so you stop generating images that have no place in the film.

1. Plan the film as timed beats

How to get it

The Herder (4:27): Jaime wrote the whole film in a note on the canvas as 11 timed beats, from the title card to a quiet ending on the clouds. Most beats name the shots they need, so the note doubles as a shot list.

To My Grandma (1:38): Zoey started from a short letter written in the first person. The narration, voiced in 10 ElevenLabs v3 clips, then directed the pictures: the video prompts are often plain first-person lines, like the letter.

The Little Fox’s Dream (1:31): The idea was one paragraph describing a single long take, and the same text went into Midjourney. One forest image from it became the film's most reused reference, in 13 generations.

Model: Text note (The Herder) · Note: One of the 11 beats

First a hand reaches through the cloud top, then his head emerges, then he swings a leg over and climbs onto the cloud surface.

Model: Seedance 2.5 (To My Grandma) · Note: A first-person line, used as the video prompt

My grandma would walk around, tending to her garden.

Watch for: A list of good-looking shots where nothing changes. Flick's Academy asks for three beats per scene (Before, Turn, After), each one a visible action, and one location with a small cast for a first film.

Step 2: Pick one look and repeat it in every prompt

The look is the film's visual style: medium, texture, light and palette. Decide it before you design characters, because every sheet and frame has to be made in it. Our guide to animation styles for AI has 15 looks with the prompts from Flick films (linked at the end).

2. Write the look once, then reuse it word for word

How to get it

The Herder: Everything is shot like a miniature model: tilt-shift, shallow depth of field. The same look is described in 196 of the 254 video prompts. Jaime picked it partly because small AI inconsistencies get lost in the blur.

Embedding Valley: A construction-paper cutout look is named in 193 of the 245 kept prompts, and one model, GPT Image 2, made 199 of the 242 images, so the paper texture never shifts between sets.

In Our Time: Here the look came from drawings, not words: 59 of the 66 stills that were animated are uploaded illustrations, many in a two-tone woodcut style, and the video prompts restate it.

Model: Kling 3.0 Pro (The Herder) · Note: Opening of a video prompt

Miniature diorama / tilt-shift macro cinematic look, shallow DOF. Heavy rain with wind-driven diagonal streaks. …

Watch for: Vague style words. A rule from Jaime's prompt cheat sheet is to "replace vague style words with visible instructions": name the texture, light and palette you want to see instead of writing "cinematic".

Step 3: Make a reference sheet for every character, set and key prop

A reference sheet (also called a turnaround or three-view sheet) shows a character from the front, side and back. Every later frame and clip is generated from it, which makes it the main way to keep a character the same from shot to shot.

3. Build the sheets once, then use them everywhere

How to get it

How much they get reused: In Embedding Valley the four character sheets were references in 118, 112, 102 and 62 generations. In A Proper Introduction the knight's sheet was used in 167, more than any other image in the film.

Props and creatures too: A Proper Introduction gave a cardboard box its own sheet, with every side and notes on size and material. In ’twixt Vows' creature head sheet was a reference in 92 generations and the vine-wrapped weapon in 59.

Start from footage that works: Sonny's worth made its sheets from frames of early test clips rather than new designs, so the sheets match what the video model had already produced.

Try two models, keep the better: In the tutorial, Jaime saves Flick's Image Generation Prompts Guide as a standing rule for the agent, then generates each reference with Nano Banana Pro and GPT Image 2.5 and keeps the better one.

Extend, don't redraw: Ulysses began with a three-view face sheet, then extended it to waist-up and full body with the face kept identical. In ’twixt Vows did the same for Lady Caerwyn:

Model: Nano Banana Pro (In ’twixt Vows)

… Image A is the approved final head-and-shoulders three-view design. … Do only one thing: extend Lady Caerwyn into a full-body three-view sheet without changing any approved design already present in Image A…

Step 4: Start with a prompt, then block the shot in 3D when the prompt can't get it

Blocking means placing the actors and the camera before you shoot. Prompts are weakest at exact camera paths and positions, and every miss is another video take at about 344 to 550 credits (cost table below). In Flick's 3D Stage you block with colored stand-ins in a rough copy of your set, record the camera move, and give that recording to the video model to follow. Parts 4.1, 4.3 and 4.4 follow Jaime Yao's tutorial above, on a 1943 scene: a wounded resistance fighter in a basement, a doctor beside him, a pianist upstairs and a patrol soldier on the street. Part 4.2 comes from Flick's 3D Stage docs and Simay Ezer Pekcan's Happy Souls Alive shot.

4.1. Previsualization without 3D skills

What it is

Previsualization: Previs is a rough version of a shot that settles where the camera goes and who stands where. In AI filmmaking you do it before you spend credits on video.

What other films did: Sketches and plain layouts. A Proper Introduction blocked shots in Blender, then redrew each as a sketch, about 70 in all. Ulysses started 158 of its 349 image prompts from a sketch. In ’twixt Vows made plain white-model layouts in GPT Image 2. 3D Stage gives the same control with no modeling or drawing.

How to get it

Build the set from one image: Select your environment image on the canvas and click To 3D Stage. The agent builds a rough 3D set from it, usually in 2 to 3 minutes, Jaime says.

Add stand-ins by asking: Ask the agent for one simple figure per character, each in its own color (Jaime's request is below). Flick also has mannequins you can pose when the body position matters.

Add the camera: Add a camera and judge the shot from its preview window, not the editor: the recording is exactly what the camera shows.

Model: Flick agent, in 3D Stage · Note: Typed as shown in the video; one color per character

Use a sphere and a cube as a character's head and body. add four characters into the scene. Yellow one lying on the bed in the basement, the red one is beside the bed. The blue one is by the piano at the first floor, and the green one is outside the building on the street.

4.2. Make a 3D animatic to time the scene

Film: Happy Souls Alive by Simay Ezer Pekcan (Flick Residency)

What it is

Animatic: An animatic is the film's shots cut together at their real length before the final pictures exist. In a 3D animatic those shots are recorded takes of your blocking, so you check the camera moves and the timing together.

Why before rendering: A move that runs long or starts early wastes a whole video take. Of the orbit in this Happy Souls Alive shot, Simay Ezer Pekcan says: "Getting it right could take countless rerolls."

How to get it

Time each move: Every camera and character gets a track on the 3D Stage timeline, and each move is a segment with its own duration or speed. Press Rehearse to play the scene from the camera's view and adjust until the move lands with the action. Simay set her orbit's duration, direction and distance this way, and previewed it a few times before recording.

Or draw it in time: On a 3D scene card on the canvas, the scene plays while you draw a camera move, and the move keeps the speed you drew it at, pauses included.

Record every camera: Record, at the top of the stage, records all cameras, each during its own segment of the timeline. The takes land on the canvas.

Cut the takes together: Select a take and use Add to Timeline to send it to Flick's video editor. In order, the takes are your animatic: the scene's length and cuts, checked before you generate any video.

Watch for: Two moves in one take. Flick's docs advise one job per camera per shot and one action per beat (walk, stop, then turn), because a combined move is hard to match with a generated clip.

4.3. Film it with your phone for real shaky cam

What it is

Shaky cam: Shaky cam is handheld footage where the frame bobs and drifts with the operator's steps, so a scene feels present and tense. In a prompt you have to describe every small movement; with your phone inside the 3D scene, you perform it.

How to get it

Connect your phone: Select the camera, drag the Manual action onto its track, click Start filming in its properties, choose Phone and scan the QR code.

Walk the shot: What you frame on the phone is what the 3D Stage camera records, shake included. Do a few practice runs, then tap record on the phone. Once you get the hang of it, Jaime says, you get "very natural handheld camera movement."

Bring it back: The take appears on the 3D Stage timeline. Click Record to capture it, then send the footage you want to the canvas as your blocking clip.

Can't film? Set handheld shake on the 3D Stage camera in its properties, or write the shake into the prompt, as In ’twixt Vows did for a walking shot:

Model: Seedance 2.0 (In ’twixt Vows) · Note: Reference to video; trims marked …

… Camera: one continuous handheld shot. … The follow is imperfect and human, with clear handheld presence throughout: visible footstep bob, irregular body sway, small balance corrections, and slight reframing as the operator walks through wet uneven ground. The camera movement is steady in intention but physically unstable from walking. … No cuts, zooms, push-ins, or pull-outs. …

Words that do the work: "visible footstep bob, irregular body sway, small balance corrections" set the shake; "steady in intention" keeps the follow readable.

4.4. From blocking clip to final shot

The rule

Each reference gets one job: Your images control the look: faces, costumes, set and light. The blocking clip controls only the camera path, the timing and where everyone stands. As Simay puts it, "The 3D video is only for camera movement."

How to get it

Brief the agent: Jaime first pastes Flick's two Seedance prompt guides into notes and tells the agent they are its rules for Seedance prompts, then saves them as project memory (Flick's docs call these Chat rules). He tells it what the blocking clip controls and who each color is (first box), and it writes the Seedance 2.5 prompt (second box).

Generate: Paste the prompt into Omni Reference with the blocking clip, the character sheets and the set. Jaime's 30-second take at 720p cost 1,440 credits, about 48 a second.

Or start from a finished frame: Simay used Extract Frame on her 3D take, replaced the stand-ins with her characters to make a painted first frame, then gave Seedance that frame, the character sheet and the 3D move. Flick's docs suggest the same: the frame guides the opening, the recording guides the motion.

Model: Flick agent · Note: Typed as shown in the video; trims marked …

I'll be using these assets as references in our omni seedance 2.5 videos. the 3d blocking video controlls the camera movements and shots composition, we also have our environment and characters references. Specify the yellow proxy represents the resistance wounded soldier, the red represents the doctor, the blue for pianist, and the green for patrol soldier. the doctor is doing a surgery to the wounded soldier, … pianist is playing piano like crazy trying to cover up the sound, the patrol soldier … loads his weapon in the end. Give me a seedance 2.5 prompt…

Model: Seedance 2.5, Omni Reference · Note: 720p, 30 s; written by the agent

For all four characters, identity, facial structure, hairstyle, body proportions, outfit, and all visible wounds, stains, and equipment 100% match their respective image references.

… reference only the blocking. Camera trajectory, framing, height, angle, roll, recorded handheld movement, timing, architectural layout, occlusion, character locations, relative scale, and base facing directions remain 100% unchanged. Add the articulated performances below within these established positions.

Live-action appearance, natural skin texture, subtle lens softness, and fine photographic grain.

…

If the agent misses a rule: Point out exactly what it got wrong, so it corrects it and doesn't repeat it. That's why Jaime asks you to read the guides yourself.

Watch for: Turning a 3D screenshot straight into a photo. Jaime found Blender screenshots came out looking synthetic, which is why A Proper Introduction went through sketches first.

When to use it: When the camera path or the positions matter and prompts keep missing. Jaime writes that he always tries prompting first; in In ’twixt Vows he took one shot to 3D Stage because it needed the camera to turn exactly 90 degrees.

682 people recorded 1,833 takes in 3D Stage (Flick data, last 90 days, October 2026, Flick staff excluded).

Step 5: Animate finished stills and give every reference a job

Most shots in these films start from a finished still or a set of reference images, not from text alone. The video prompt then does three jobs: it says what each reference controls, sets the camera, and describes the action in order.

5.1. Say what each reference controls

How to get it

In ’twixt Vows: 183 of the 263 kept video prompts tie each character to their sheet with an "appearance match" line, and 97 also pass a voice clip for the character's voice and accent. 130 number their shots ("Shot 1") so Seedance can cut inside one clip.

Embedding Valley: 62 of the 86 kept video prompts pass a voice clip, 76 say exactly how many shots to make, and 45 ask for hard cuts with no dissolves.

Model: Seedance 2.0 (Embedding Valley) · Note: @ names a canvas item; the sheet's Chinese name means three-view character sheet

Congxing's appearance match @Congxing 三视图角色设定 100%.
Reference Congxing's voice and accent from @Usage Growth Projection.
…
Exactly 2 shots in total.
Exactly 1 hard cut.
Do not create any additional shots.
Do not morph, dissolve, or add transitions between the shots.
…

5.2. Set the camera first, then one action at a time

How to get it

A Proper Introduction: 191 of the 206 kept video prompts open with the camera, handheld or locked, then describe the action one line at a time. 163 ask for no background music, leaving sound for the edit.

In Our Time: 62 of the 98 video prompts start with a locked camera, then name the one thing that may move.

Joining shots: For 44 of The Herder's clips the model got a first and a last frame and filled in the motion between them. The Little Fox’s Dream pulled 20 frames from finished clips so the next shot starts where the last one ended.

Model: Seedance Pro (In Our Time)

Locked camera, same composition and style as the source image: woodcut-like linework, strict two-tone palette of deep desaturated blue and off-white, no new details, no change in lighting, no extra textures.
Only animate the salmon and the waterfall inside the doorway; the doorway, trench, sky and foreground stay completely still.
…

Model: Seedance 2.5 (Sonny's worth) · Note: The retake that worked

The woman screams hysterically: "Do it!"
when she says "it!", the camera quickly zoom pans into her mouth until the screen is dark
don't fade to black, the scene should be at full opacity the entire duration, it just zooms into the darkness of her mouth
don't start the zoom in until she says "it", before that camera should be locked

Watch for: A move that starts too early, or a fade nobody asked for. The scream in Sonny's worth was first tried as a video edit, then regenerated with the exact moment the zoom may start and what must not happen (second box).

Step 6: Cut the film and add sound in Flick

The breakdowns say less about the edit than the earlier steps. What they show: voices and music were generated next to the pictures, and of the 9, only the Sonny's worth breakdown describes the cut: it was made in Flick's video editor.

• Record the voices before the video when there is dialogue. Embedding Valley generated 15 ElevenLabs v3 voice clips and passed them to Seedance as voice references. Its score is 12 cues made with Eleven Music.
• Captions and the cut. Sonny's worth made its dialogue with ElevenLabs v3 and captions with ElevenLabs Scribe, and was cut in Flick's video editor with trims, clip effects, adjustment layers and title text.
• Make graphics in the film's look. Embedding Valley's name cards were generated as paper cutouts too, so the titles belong to the same world.
• Cut early and without mercy. Flick's Academy suggests assembling a rough cut of the whole film early, and dropping a shot that doesn't help the story even if it was hard or expensive to make.

Which AI models make movies, and what one take costs

The 9 films lean on a few models. Nano Banana Pro is the main image model in 5 of them; in the other 4 it's Nano Banana (2 films), GPT Image 2 or Midjourney. For motion, 8 of the 9 used Seedance, and Kling was the main video model in 3.

Model	Typical credits	Typical time
GPT Image 2	about 9	about 75 s
GPT Image 2.5	about 9	about 33 s
Nano Banana Pro	about 32	about 45 s
Seedance 2.0 Mini with references	about 77	about 3.5 min
Seedance 2.0 with references	about 344	about 5 min
Seedance 2.5 with references	about 550	about 6 min

Flick data, successful generations in the last 90 days, October 2026. Averages; a longer or higher-resolution take costs more.

Rough math for one finished minute: the median film made about 26 video generations per finished minute. At about 344 credits a take with Seedance 2.0 and references, that is close to 9,000 credits of video before any images. Treat it as an order of magnitude, since take lengths vary and several of these films used cheaper or older models.

Make your first AI movie in Flick

Sign up, open a canvas and block your first scene in 3D before you generate.

Every account gets 100 free credits when you sign up, and 100 more every Monday.

Create your free Flick account →

Keep going

• Animation styles for AI 15 looks, each with the real prompt from a film made in Flick.
• AI camera movements 12 camera moves with prompt wording, and when to draw a move in 3D instead.
• Behind In ’twixt Vows Jaime Yao on real-world references, which model does what, and 3D Stage camera work.
• How I made a Viking film with AI Jaime Yao's full Ulysses workflow, from sketches to Kling dialogue takes.
• Behind The Herder How the miniature look and the turnaround sheets came about.
• Make your first film (Flick Academy) Flick's step-by-step course, from story brief to finished cut.
• Image Generation Prompts Guide (docs) The guide Jaime saves as the agent's rules for still images.
• Seedance Performance Guide (docs)
• Seedance Action Guide (docs)
• 3D Stage basics (docs) Stand-ins, props and cameras; the agent can build the scene from your image.
• 3D Stage cameras (docs) Lens, framing and every camera action, including Manual.
• 3D Stage export (docs) Record one camera or all of them, then use the take in Omni Reference.
• Omni Reference (docs) Which reference controls what: start frame, character, motion and voice.
• Talking to the agent (docs) How to brief the agent, and where Chat rules live.

About the author

Bilguun Bayarkhuu is a growth engineer at Flick. This guide is built from the public workflow breakdowns of 9 films made in Flick by Jaime Yao and Zoey Z., from Jaime's 3D tutorial on Flick's YouTube channel, and from Simay Ezer Pekcan's Happy Souls Alive tutorial.

Frequently Asked Questions

Can I make a movie using AI?

Yes, and alone. Ulysses, a 4:36 short film, was made by one person with no crew, and To My Grandma (1:38) needed only 22 images and 13 video clips. The work is in the planning: beats, one look, reference sheets and a plan for each shot, then video.

How do they make AI films?

Most of the films above follow this order: write the beats, fix one look, make reference sheets for the main characters and sets, plan each shot with a sketch or by blocking it in 3D, turn the plan into finished stills, animate the stills with a video model such as Seedance or Kling, then edit with voices, music and captions. About 72% of all generations were stills.

How long does it take to make an AI movie?

Jaime Yao made Ulysses (4:36) in about 8 days, running 432 generation jobs, and he says most of that time went into choosing between takes rather than waiting on renders. In Flick a video take typically takes 3.5 to 6 minutes to generate, and an image under a minute and a half.

How much does it cost to make an AI movie?

In Flick it's paid in credits, and video is most of it: about 344 to 550 credits for a typical Seedance 2.0 or 2.5 take with references, against about 9 to 32 for an image. The 9 films above made a median of about 26 video generations per finished minute. Flick's monthly plans include 500 (Casual), 2,500 (Standard), 8,000 (Pro) or 24,000 (Studio) credits.