Announcing Scene Flow: AI Assisted Scene Creation

Updated 2026-10-11·15 min read·Announcements

The screenplay and the storyboard on one page. We took a one-paragraph premise to a voiced, rendered six-shot dialogue scene without leaving it. Every image below is a screenshot from that session.

Scene 4, INT. WHITFIELD HOME — KITCHEN — DAWN, as it stands at the end of this walkthrough. 9:16, photoreal, 24 seconds of picture. Demo project: Her Stolen Life, Episode 1, built October 2026 on showmaker.studio.

Most AI video tools give you two places to work. You write in one, then carry the scene into another to make pictures, and the two drift apart the moment you change a line. Scene Flow puts them on the same page. The scene's action beats and dialogue lines run down the middle of the screen as a screenplay. Each shot sits under the beat or line it covers, as a card holding its start frame, its references and its finished still. A preview of the whole scene plays alongside.

Edit a line and the shot under it knows. Draw a frame and the script knows. Nothing needs to be synced by hand.

Here is one scene, built from nothing, in the order we did it.

Step 1 · It starts as one paragraph

Our premise was a classic "husband stealer" setup, placed deliberately in suburban Atlanta. Dana Whitfield is a pediatric ER nurse who works nights. Her husband Marcus owns a struggling car dealership. Her best friend from college, Kelsey Brooks, moves into the guest room after a divorce and starts quietly taking Dana's place: cooking Marcus's breakfast, wearing Dana's robe, sitting in Dana's seat at church.

We pasted it into the Drama Builder and pressed Generate 60-episode outline.

The Drama Builder. The premise box, an episode count (12 to 60), the audience and market, and an optional trope engine. One press writes the whole series outline.

The outline came back in 74 seconds as Her Stolen Life. It names Dana's wound ("feeling invisible and taken for granted in her own home") and lays out four ladders she climbs over 60 episodes. It also scores the outline against the patterns that make short dramas travel: thumb-stop open, dramatic irony, status reversals, cliffhanger discipline. This one scored 94 out of 100.

The outline card. Six of the ten scores are computed from the outline's structure and four are the model's own assessment; the card says which is which. Kelsey is marked as the recurring villain, with her return episodes listed.

Locking the outline in creates a series project and one project per episode. In the background, Episode 1 gets a written screenplay of seven scenes with action beats and dialogue. The cast (Dana, Marcus, Kelsey and a friend named Brenda) and six locations are created with it, along with each character's outfits, such as Dana's nurse scrubs and Kelsey's real-estate-agent look.

Step 2 · The scene opens as a screenplay

Scene 4 is the morning Dana notices. She comes home from a night shift to find Kelsey in her kitchen, in her robe, and Marcus not looking up from his phone. In Scene Flow it opened like this:

Scene 4, untouched. The bar across the top reads “9 beats · 3 lines · no shots yet”, the honest state of the scene. Underneath are the master frame, its audit, and the action text. The preview on the right already plays the scene's one establishing frame.

Three things on this screen are easy to miss and matter a lot.

The scene tells you what it is missing

The red BLOCKER says that three references this scene uses (@dana_whitfield, @kelsey_brooks and @whitfield_home) have no reference still. Without one, every shot would invent Dana's face and the kitchen afresh, and differently each time. The fix is on the banner: Draw all 3 · 12 cr.

The audit reads the scene before you spend anything

Below the blocker, an audit lists what will go wrong. The master is framed as a medium close-up, so it cannot establish the room. The action tags Dana but never one of her two outfits, so the model would dress her however it likes. Fix 9 issues… rewrites the prompts to clear these flags. It is text only, costs no credits, and shows you every change before anything is saved.

Everything is a tag

Characters, outfits, rooms and props are @slugs. A prompt that names @dana_whitfield_nurse_scrubs sends the drawn wardrobe sheet to the image model with it. That is how the scrubs stay the same scrubs across six shots.

Step 3 · Draw the people and the room once

One press on Draw all 3 drew a character sheet for Dana and one for Kelsey, plus a plate of the Whitfield kitchen. The button counts down while it works.

After the draw. The missing-reference flags are gone from the audit, and the master frame's Refs row now shows three thumbnails. Those are the images every later shot of this scene is drawn against.

Step 4 · Break into shots

Break into shots reads the beats and the three lines and plans the coverage the way an editor would want it: an establishing wide, a single on each speaker, an insert, and a reaction. It took 22 seconds. Each shot was filed under the beat or line it covers, and the button turned into Re-cover shots.

After the breakdown. The top bar now reads “6 shots · 6 in cut”, and the scene preview is 0:27 long. The blocker updated itself as well: the new shots name Marcus and three specific outfits, so it asked for those sheets next. We drew them the same way.

Step 5 · Frame each shot in conversation

Every shot card has a Shoot with AI button. It opens a two-step dialog: first the still (the start frame), then the video (what moves). The AI writes a prompt for each step. You can edit the prompt directly or tell the AI what to change in plain words. Nothing reaches the shot until you press Save to the shot.

Here is Shot 2, Dana's line: "That's my robe, Kelsey."

The still step, first take. At the top are the line, who speaks it and how it is voiced, and the framing. The block below it is the AI's continuity sheet. On the left is the drawn still; on the right, the prompt that drew it. The model picker shows each draw's price (Nano Banana 2 Lite · 2 cr).

The AI keeps a continuity sheet

The panel headed What the AI knows at the start of this shot is compiled from the beats and the shots before it. It records where the light comes from ("cold blue pre-dawn light raking in from the window behind Dana"), how the previous shot ends, and which side of the frame each person holds. You can correct any line of it, and a correction is treated as fact for every later turn. The toggle beside the framing, Continues the shot before / New setup, controls how much this shot inherits from the one before it.

It reads the picture, not just the prompt

The orange note under the prompt comes from the AI looking at the image already on the shot. It reported that the frame held Dana in her home outfit while the prompt asked for scrubs, and it said which one wins: "The picture is the fact." Stored prompts go stale after a re-shoot, and the picture does not.

Then we corrected it

The first draw put Dana in a cream nightgown. She has just walked in from a twelve-hour night shift, so that is wrong. Rather than rewriting the prompt by hand, we typed one note into the box:

DIRECTOR: She is still in her nurse scrubs from the night shift — keep her in the scrubs, not a nightgown.

Left: the note goes in the chat box, which suggests examples like "closer", "he's facing camera" or "the room is dark". Right: the AI rewrote the prompt and we redrew: navy scrubs, three-quarter profile, the window light behind her carving an edge down her jaw. Use the new still puts it on the shot.

Then the motion

Step two writes the video prompt from the frame you just approved. It keeps the still's framing and cast exactly, adds the camera behaviour, the light, where focus sits and the performance, and puts the line in verbatim. The AI explains its choices in a sentence above the prompt. Here it kept the camera locked off "so the whole five seconds rests on her face".

The video step. The start frame stays on screen while you work on the motion. Save to the shot writes the still and both prompts at once.

We ran the other five shots the same way, each with one continuity note about the wardrobe. Each took under a minute.

Step 6 · Every line with its shot under it

With the frames drawn, the scene reads top to bottom as script and storyboard together. Each dialogue line shows who says it, an estimate of its spoken length, and a Generate voice button. Directly below it is the shot that covers it.

A shot card. From left: the start frame (with an end frame slot for a morph), Use as master, the reference sheets this shot draws against, and the finished still with its length. Along the bottom are the render model (here the free Ken Burns animatic), the voicing and the timing. Render Options picks a paid video model when you are ready.

Left: Kelsey's line and its shot. ~10.9s spoken sizes the shot before a voice exists. Speaks marks the shot that performs the line on camera. Right: Marcus's line. The red bar says the line is not voiced yet, and names it word for word. Render scene in Step 8 records it.

Step 7 · Simple or Cinematic

The top bar has a switch that decides how Scene Flow writes video prompts for the whole project.

• Simple — the default. A compact prompt that still decides shot size, camera height, light, focus and sound. It suits a fast-moving microdrama where the performance carries the scene.
• Cinematic — longer prose in the style of our Board Agent. The camera moves for a stated reason, delivery verbs are specific ("says flatly", "says warmly"), and each shot ends where the cut should leave the eye.

The choice is saved with the project, so everyone working on it gets the same writer. It applies to Shoot with AI and to every Refresh of a video prompt.

Step 8 · Render Options, then Render scene

Every shot card has a Render Options button. It opens the shot's settings in a panel beside the scene, so you can change a shot and keep the Render scene button in view. Here is Shot 2, Dana's line, top to bottom.

Left, the upper half: what kind of shot it is and how its line is spoken. Right, the lower half: sound, captions, finishing touches, and the shot's own still prompt.

What each setting does

The Model menu prints each model's price and abilities in the menu itself, so the cost of a choice is visible before you make it.

Then we pressed Render scene

We left every shot on the free animatic and pressed Render scene. For each shot whose picture or sound changed, it records the voice, renders the Ken Burns clip, burns in the captions and mixes the sound. Then it assembles the six shots into one clip. A shot that hasn't changed since its last render is left alone, so a second press costs nothing for it. The button reports progress as it goes: Rendering 3 shots… 4 left, then Assembling scene….

Left: while it works, each shot card shows its own Rendering badge, and the button counts down. Right: the result, exactly as Render scene produced it: 24 seconds, 1080×1920, three voiced lines over six animated stills, with captions. Play the rendered scene (24 s)

After the render. Dana's line now has a playable take and a Regenerate button, Copy voiceover has appeared at the top of the options, and the preview's Video tab plays the assembled scene. To upgrade a shot later, change its Model to a video model and press Render scene again; only that shot re-renders.

What we never did

We never opened the Script tab, never opened the Storyboard, and never copied a prompt from one place to another. The premise became an outline. The outline became a screenplay, the screenplay became six framed shots on one page, and one press turned those into a voiced, captioned 24-second scene. Each step showed its cost before we pressed it. The frames ran 2 credits a draw on Nano Banana 2 Lite, and the render itself was free on the animatic.

Screenshots: showmaker.studio, production, 11 October 2026. Demo project Her Stolen Life (Episode 1, Scene 4) was created for this article from the premise quoted above. Nothing has been retouched; images are cropped and resized only.