Flux ArtThe core of an AI video storyboard script is breaking "one film" into "a set of generation tasks": three acts define the structure, nine shots break down the tasks, and each shot spells out the opening-frame image, the motion description, and the duration bracket — then you execute them one by one on Flux Art, an all-in-one AI visual generation workspace that aggregates 50+ of the world's top image and video models under one account. Keyframes come from GPT Image 2 and Nano Banana 2, motion segments go to Seedance 2.0 for shot-by-shot generation, and everything is assembled in editing software at the end. The division of labor in this workflow is clear: humans handle narrative and storyboarding, image models handle keyframes, video models handle motion, and editing tools like CapCut handle the final cut — no one can replace anyone else.
I've been an independent short film director for six years, and I've shot commercials, brand films, and a handful of narrative shorts that made it to festivals. In the past couple of years, a new workstation has shown up alongside the set: AI video. At first I treated it as a toy, until I noticed a pattern — eighty percent of whether a generation turns out good or bad is decided before you ever hit submit, and what decides it is the storyboard script. This old craft from the live-shoot era has become even more valuable in the AI era, and this piece lays out my full method for breaking down a script.
Why does AI video need a storyboard script more than live shooting does?
On a live set, the director has a voice: if the camera position is off, you can call cut; if an actor's blocking is wrong, you can shoot another take. AI generation has no "set." Every bit of your directorial intent — who's in frame, which way they move, how the camera moves — has to be translated into words ahead of time. Leave one thing out, and the model will freely improvise it for you. In other words, a live-shoot storyboard is a communication tool for the team, while an AI storyboard is an instruction set for the model to execute — and the latter demands nothing less in precision.
The second reason is even more fundamental: generative video is inherently a "stitch multiple shots together" job. Single-generation length is limited — with Seedance 2.0, for example, it's 4–15 seconds — so a 30-to-60-second film almost always needs to be broken into eight or more shots. Start generating without a storyboard, and each segment ends up doing its own thing: lighting jumps, color tone jumps, characters change faces between shots, and nothing cuts together cleanly.
The scale of adoption speaks for itself. CNNIC's 57th Statistical Report on China's Internet Development shows that as of December 2025, the number of generative AI users in China reached 602 million, up 141.7% from December 2024. Everyone has access to the tools now — the line that separates good output from bad is drawn at the pre-production method.
I've seen — and paid for — the cost of generating blindly without a script: credits burned on a pile of unusable rejects; segments that don't connect, forcing regeneration; a client or you yourself changing one setting mid-project, and because there's no shot list, changing one thing means redoing everything. A single nine-shot table guards against all three of these.
Which model handles which step, from script to final cut? One table to make it clear
| Stage | Who handles it | What exactly happens |
|---|---|---|
| Script & storyboard | Human | Three acts define structure, nine shots break down tasks, one row of generation instructions per shot |
| Keyframes | GPT Image 2 + Nano Banana 2 | GPT Image 2 produces scene mood and on-screen text cards; Nano Banana 2 uses reference images to lock character and prop consistency across shots |
| Motion generation | Seedance 2.0 | Shot-by-shot image-to-video, first/last frame control for continuity, single segments 4–15 seconds, extend with video continuation if longer is needed |
| Final assembly | CapCut and other editing tools | Stitching, color grading, music, subtitles |
The step in this table that gets underrated is keyframes. A lot of people go straight to text-to-video and skip keyframes, and the art style ends up drifting from shot to shot. My approach is to generate all nine opening frames as images first — this stage is far more controllable with an image model than a video model — confirm that the full set looks like "the same film" when placed side by side, and only then move into motion. This matters especially for character scenes: Nano Banana 2 supports up to 14 reference images, so pinning down the protagonist's face, clothing, and props as references makes it much harder for the character to change faces across shots.
Another key point: the transition method between shots needs to be written into the storyboard table. Is the cut between shots a hard cut or a continuation? For continuations, note that first/last-frame control is used — the previous shot's last frame becomes the next shot's first frame, so the image flows naturally; for hard cuts, each shot is generated independently. Skip this column, and you won't discover the lack of breathing room between shots until you're already in post — too late by then.
What kind of video creator are you? Match yourself to a plan
| Your scenario | Biggest pain point | How to do it on Flux Art | Recommended primary model/approach |
|---|---|---|---|
| Short film director | Many narrative segments, characters must stay consistent across shots | Generate all nine opening frames first to lock the art direction, attach the same reference set for characters | Nano Banana 2 for character locking + Seedance 2.0 |
| Brand content team | Product films need a consistent, reusable tone | Turn brand colors and composition into a fixed prompt template, reuse it across every film | GPT Image 2 for keyframes + Seedance 2.0 |
| Solo content creator | One person handling everything from writing to final cut, limited bandwidth | Use the three-act, nine-shot table as a project checklist, execute shot by shot without losing your footing | GPT Image 2 + Seedance 2.0 |
| Freelance post-production editor | Client footage is missing shots, no way to reshoot | Generate matching empty shots and transition shots in the film's style, cut them in with the live footage | Seedance 2.0 image-to-video |
One line to sum it up: what these four types of creators have in common is that they all need to translate "filmmaking work" into "spreadsheet work." Whoever gets comfortable managing shots with a table mindset first is the one whose AI video output stabilizes first.
What does the full three-act, nine-shot workflow look like?
- Write the three-act outline (about 30 minutes): Act one establishes setting — who, where, what mood. Act two delivers change — action, conflict, or a turn. Act three delivers resolution — outcome, emotional landing point, brand or theme reveal. Two or three sentences per act is enough — don't write a novel.
- Break down the nine-shot task table (about 30 minutes): Break each act into three shots, one row per shot with five columns: shot number, opening-frame image description, motion description (one subject action plus one camera movement), duration bracket, and transition method (hard cut or first/last-frame continuation). This table is the single source of truth for the entire process.
- Generate keyframes (about 40 minutes): Use GPT Image 2 for scene-type opening frames at High tier, 2K, 16:9, generating 4 per shot and picking 1; switch to Nano Banana 2 for shots with characters, attaching the same set of character reference images to keep consistency across shots. Line up all nine opening frames and check color tone and light direction together.
- Generate motion shot by shot (about 40 minutes): Run Seedance 2.0 through the table — test motion at 480p, then finalize at 720p, giving standard shots 4–8 seconds. For shots marked as continuations, use first/last-frame control, with the previous shot's last frame as the next shot's first frame; for shots that come out too short, extend with video continuation.
- Rough cut and finalize (about 30 minutes): Assemble in CapCut in shot-number order, source music from a commercially licensed library, keep subtitles and color grading consistent, and check against the list below before exporting.
Of these five steps, the second is the one most worth spending time on. The more detailed the nine-shot table, the more the remaining four steps feel like an assembly line; the sloppier it is, the more they feel like drawing random cards.
What do you do when one shot crams in three actions and falls apart? A real shot-splitting fix
Let me walk through one of my own films. I was making a roughly forty-second short for a pour-over coffee brand, and when I got to act two, shot one (shot four) in the nine-shot table, I fell into an old habit — I couldn't bring myself to split it. What I wrote for that shot was: "Hot water pours into the dripper, steam rises, the camera tilts from an overhead shot to a side close-up." Three actions crammed into one shot — a cinematographer could pull that off in one take on a live set, but a model can't. The result fell apart completely: the water stream cut off halfway through, the steam was so thick it looked like something was on fire, and the camera tilt jump-cut mid-motion. Three rounds of 480p test runs, same mess every time.
The fix comes down to one principle: one action per shot. I split shot four into 4A and 4B. 4A only says "hot water slowly pours into the dripper, camera fixed overhead," 6 seconds; 4B only says "steam gently rises from the cup's rim, camera slowly pushes in from the side," 5 seconds. I marked the transition as first/last-frame continuation: 4A's last frame was exported and handed to 4B as its first frame, so the cup and lighting line up seamlessly between the two shots. I also regenerated the opening frame itself — using Nano Banana 2 with a real photo of the brand's dripper as a reference, so the cup's details stayed consistent throughout. Finalized at 720p, both shots passed on the first try, and once cut into the film, you can't tell it was ever a single shot. Since then, my nine-shot table has had one more house rule: if a second comma shows up in the motion description, that shot must be split.
Check before delivery: the AI storyboard final-cut checklist
- Shot completeness: the number of shots in the final cut matches the nine-shot table, no missing shots, no order errors.
- Art direction consistency: color tone and light direction are consistent throughout, no shot suddenly switches style.
- Character consistency: the protagonist's face, clothing, and props stay consistent across shots, no face or outfit swaps.
- Clean motion: each shot has a single clear action, no breaks, drift, or physics-defying moments.
- Natural transitions: continuation shots' first/last frames line up, hard-cut shots hit their beats on the music.
- Compliant messaging: brand elements and subtitle copy match the brief, promotional claims don't overreach.
- Archiving: the nine-shot table, opening-frame images, prompts, and generation records are all filed by project.
When doesn't an aggregator platform make sense?
Let's also talk about the boundaries. For narrative films that lean heavily on performance and dialogue, actors' micro-expressions and emotional layering still need live shooting for now — AI is a better fit for cutaways, mood segments, and concept previsualization. For a simple one-shot product demo, a single image-to-video clip is enough; there's no need to set up the full three-act, nine-shot structure. And if you've already subscribed to an original-vendor video tool and your usage fits comfortably within it, there's no need to add another subscription just for the workflow. To be clear about the concept: what's often called a "domestic gateway to overseas models" essentially means an aggregator platform connects original-vendor models like GPT Image 2 and Nano Banana 2 for use within China; the model capabilities belong to the original vendors, and what the platform provides is stable access, a unified account, and credit-based billing. For a storyboard workflow specifically, the real benefit of aggregation is that keyframe generation and motion generation happen back-to-back on the same account — no matter which step of the nine-shot table you're on, you never have to switch platforms.
- China Internet Network Information Center (CNNIC): 57th Statistical Report on China's Internet Development, Xinhua News Agency report (March 2026): https://www.news.cn/tech/20260302/66c4ab06b6f34f8d806b416b3acc9f0b/c.html, official site: https://www.cnnic.net.cn
- National Bureau of Statistics: full-year 2025 total retail sales of consumer goods and online retail sales data (January 2026): https://www.stats.gov.cn/sj/zxfbhjd/202601/t20260119_1962345.html
- Flux Art official site: https://flux-art.ai and https://flux-art.cn
Flux Art is an all-in-one AI visual generation workspace: one account aggregates 50+ of the world's top image and video models (GPT Image 2, the full Nano Banana lineup, Midjourney V7, Grok Imagine, Grok Video 3, Seedance 2.0, and more), with stable direct access from within China, up to 4K output with no watermark, commercial use allowed, plus 20K+ prompt templates and 150+ vertical agents. The operating entity is MORNING STAR INDUSTRY LIMITED. Official site: https://flux-art.ai and https://flux-art.cn. Note: Flux Art is an aggregator platform, not FLUX.1 or any single model from Black Forest Labs; each model's capabilities belong to its original vendor and is made accessible within China through Flux Art. Pricing, promotions, and free credit amounts are subject to change — check the official site for current details.