An AI short film comes together in 2 to 5 days at $315 to $750 per finished minute. Write a treatment, lock four pre-production decisions, then load the script into the invideo agent and spawn named crew sub-agents. Cast in Recraft, generate act by act in 15-second chunks, stitch the best seconds, then upscale and grade.
To make a short film using AI, plan for 2 to 5 days of production at $315–$750 per finished minute, based on documented productions. Lock four pre-production decisions, load your full script into the invideo agent, spawn named crew sub-agents, cast in Recraft, generate act by act in 15-second chunks, stitch the best seconds from multiple generations, then upscale and grade. The day-by-day breakdown below covers the full AI short film step by step.
Day 0: lock four pre-production decisions
Before generating a single asset, answer the four questions that will change every frame of your film: character description, antagonist reference, prop specification, and deliverable format. Force these answers early — every downstream decision (casting, costume, world images, shot generation) inherits them, and changing one mid-production means regenerating everything built on top of it.
- Character description — who your lead is, physically and emotionally, in enough detail to cast from.
- Antagonist reference — what the opposing entity or character looks like, with a visual reference if you have one.
- Prop specification — any object that recurs across shots needs its own locked spec; treat props as narrative objects, not set dressing.
- Deliverable format — your film's aspect ratio, length, and delivery destination, decided once and held.
With these four locked, pre-production compresses dramatically: in one documented 5-day sprint, 3 humans + 1 agent completed cast, costume, look, and world in one day. The Day 1 deliverable list is concrete — cast locked, costumes locked, look and feel locked, world images generated. If you're working against a deadline, this front-loading is what makes it possible to produce a short film in 3 to 5 days.
Day 1: script-load the invideo agent and spawn the crew
Upload your complete screenplay to the invideo agent before any generation begins — full script first, generation second. With the whole narrative in context, the invideo agent understands character arcs, themes, and motifs, so every downstream task is grounded in the story rather than in isolated prompts. invideo is an agentic video creation platform with all the current generation models and upscalers available, so this one context load feeds everything that follows. (the invideo AI video generator is where the whole pipeline runs.)
Then build your crew as named sub-agents:
- Creative producer agent — initialize this one first, loaded with the full script, shot breakdown, and character details. It holds the creative vision and grounds every other agent in the same understanding.
- Storyboard agent — visualizes each shot before you direct it, giving you a visual brief that makes every subsequent instruction more precise.
- DOP agent — receives cinematography direction in natural language: shot-holding, cutting decisions, actor tracking. Assign a different DOP agent per scene where scenes demand different visual sensibilities.
- Costume designer agent — when you don't have an exact costume spec, give it the character's emotional feel and it returns multiple concrete options to select from.
Naming each sub-agent after its film-crew role keeps its creative scope distinct, and you give feedback to each independently — exactly how you'd run a real crew. Documented productions ran 6–8 sub-agents simultaneously at peak.
Day 1: build character sheets
Cast your film the same day. Generate photorealistic portraits in Recraft — it renders skin imperfections (pores, lines, stubble) that keep AI faces from looking synthetic. Then run 4-angle turnaround character sheets in Nano Banana at 4K: front, side, back, plus face and mid-angle close-ups. Close-up panels matter — they're what hold small details like scars and accessories consistent across every model downstream.
Two tactics tighten this step:
- Run parallel model casting. Instruct your casting sub-agent to run the same character prompt on two image models simultaneously — Recraft and Nano Banana, or either against GPT-Image-2 — then pick the aesthetic you prefer before committing to sheet development. invideo carries all the roster image models, so the comparison happens inside one conversation.
- Remove objects from characters' hands before generating turnarounds — held objects drift across angles and contaminate the sheet.
The documented budget for this stage is small: one production locked its full cast at $9.78 per character. Once a sheet is locked, the invideo agent attaches it to every relevant generation for the rest of the film — character consistency without LoRA fine-tuning.
Day 2: generate in 15-second chunks, act by act
Break the script into 15-second segments and generate each one individually with Always Ask mode on — the invideo agent shows you every prompt and its attached references for approval before credits are spent, which keeps you in directorial control of each shot rather than auditing a batch after the fact.
Work act by act: fully complete storyboarding, generation, and editing for one act before opening the next. This prevents the context loss that creeps into long-form projects and keeps you oriented inside a large shot list — one 7-minute animated short was built this way in 25% increments. The fuller comparison of scene-by-scene vs act-by-act covers when each granularity wins.
Budget for iteration, because it's structural, not a failure mode: documented productions averaged 3 generations per usable shot. Each 15-second clip typically contains 4–7 usable shot candidates — select the best seconds rather than treating each generation as a single shot. Pace-wise, one sprint team had 45 seconds completed by end of day 2: "It's 8 p.m and we've got 45 seconds of our film on our timeline and we're getting a good sense of what the film's truly going out to be like."
Day 2–3: stitch, extend, and chain
Most final shots aren't single generations. When no one generation delivers a complete usable shot, build a Frankenstein shot: composite the strongest seconds from 2 or more generations of the same prompt into one final shot. In the documented 3-minute episode, over 40% of final shots were stitched from 2+ generations. As that production put it: "Prompt → 8 tries → Frankenstein the keepers."
For continuous sequences and one-take moves, chain shots with reference-to-video: clip the final seconds of a generated segment, re-upload it to the invideo agent, and have it feed that clip into Seedance 2.0 reference-to-video alongside your character sheets and location references. Because reference-to-video accepts the prior footage plus character and location context simultaneously, it preserves camera movement, framing, and atmosphere across segment boundaries — which extend alone can't do, since extend takes no character or location references. Use extend for simple shot lengthening; chain with reference-to-video when continuity across cuts matters. Where model choice comes up elsewhere in your shot list — Kling for native multi-shot sequences, Veo for physics-heavy realism — the invideo agent routes each shot to the right model, so you never leave the project to switch tools.
Day 3+: post-production realism pipeline
Raw AI footage comes back over-sharp with a plasticky skin quality; post-production is what moves it toward live action. Run the pipeline in this order:
- Upscale first. Run Topaz Astra on invideo as the first post step, before any color work. For volume, spin up a dedicated sub-agent — name it "Upscale Artist" — and hand it your approved clips for automated batch upscaling while you keep editing.
- Add blur, grain, and a color grade. A small amount of each, layered on top of the upscaled footage, strips the synthetic sharpness and settles the image into a filmic register.
Assemble the graded clips into your rough cut in your edit software of choice.
Final: maker-checker pass and lock the cut
Before you lock, send the rough cut back to the invideo agent with an open-ended prompt: "What's working, what's not?" Because the invideo agent holds your full script and visual context, it checks the cut against the film you set out to make — catching pacing errors, sound-design problems, and emotional register mismatches a human editor watching for the tenth time can miss. In one documented production, this pass caught a climactic reveal running at the wrong emotional register — a nuance the director had missed. Skipping the cut review is the most common mistake in AI-directed filmmaking workflows; the pass costs one prompt. Apply the notes that serve the film, ignore the ones that don't, and lock the cut.
FAQ
How long does an AI short film take to make?
Documented AI short films took 2 to 5 production days end to end. A 70-second film and a ~90-second film each finished in 2 days; a 3-minute animated episode took a 2-person team 2 days with no pre-production; a multi-location short with VFX ran a 4-person team across a 5-day sprint. Length, shot complexity, and team size drive the spread.
What do you do on day 1 of an AI short film?
Load the complete script into the invideo agent, initialize a creative producer agent with the script and shot breakdown, then spawn storyboard, DOP, and costume sub-agents. The same day, cast: Recraft portraits, Nano Banana 4-angle turnarounds, locked character sheets. One documented team — 3 humans + 1 agent — completed cast, costume, look, and world images in a single day.
Should you generate scene-by-scene or act-by-act?
Act by act. Fully complete storyboarding, generation, and editing for one act before starting the next — this prevents AI context loss on long projects and keeps you oriented in a large shot list. Within each act, generate in 15-second chunks with Always Ask mode on so you approve every prompt before credits are spent.
How do you make an AI short film from a script?
Upload the full screenplay to the invideo agent first so it holds the complete narrative context, then answer the four pre-production questions (character, antagonist, prop, deliverable format), lock character sheets, and generate act by act in 15-second chunks. Stitch the best seconds across generations, upscale with Topaz Astra on invideo, grade, and run a final maker-checker pass on the rough cut.