How to make a short film with AI: the full workflow, with real costs

A fifteen-second clip and a three-minute film are two different things, and length is the smallest of the differences.

At fifteen seconds you can get away with one generation, one prompt, one lucky take. At three minutes you have forty shots, three or four locations, at least two characters who have to look like themselves in every scene, a soundtrack, and an edit. None of those problems exist at fifteen seconds, and all of them compound: a character who drifts in shot 12 has to be fixed in shots 13 through 40 too.

This is the workflow for the three-minute version, written for someone who has made clips and wants to make a film with AI. It ends with the question everyone asks and few answer straight: what does it cost?

What counts as a short film here? Past one scene, every problem changes

For this article, a short film is: a story with a beginning, a middle and an end, told across more than one scene, running two to five minutes, with the moving images generated with AI. Thirty to sixty shots is typical.

The reason to draw the line at "more than one scene" is that scenes are where the difficulty lives. Within one scene, one location, one lighting setup, one costume, the tools are good at keeping things stable. Between scenes, everything is allowed to change, which means everything is allowed to go wrong: faces, clothes, the colour of a door, the time of day.

So the workflow below is really a set of habits for surviving scene changes. If you only need one scene, the beginner's guide to making a video with AI is the shorter route.

The film is decided before you generate: script, shot list, locked assets

This is the part people skip, and it is the part that decides the film.

In 2026, Higgsfield released a 95-minute feature made with AI, Hell Grind, made by a team of fifteen in a fourteen-day generation window on a budget under half a million dollars, and published the entire production: every prompt, every asset, every working rule. Our count of that project data found something more useful than the headline numbers. The team's prompts got longer as the shoot went on: the median prompt for scenes generated around day one was about 7,000 characters; by day fourteen it was about 32,500. Four times longer in two weeks. The later prompts had whole sections the early ones did not: explicit constraints, camera rules, sound rules. Their published "production bible" is, in effect, what they wished they had written on day one.

You do not need 32,000-character prompts for a three-minute film. You do need the same four things before you generate anything:

1. A script. Three minutes is roughly three beats of story: a situation, a turn, a consequence. Write it as prose if formatted screenplays scare you. Every scene needs a reason to exist.
2. A shot list. Every scene broken into shots, with size, angle, movement and length per shot. Script to storyboard: how to break a scene into shots walks through this with a worked example and a template.
3. A storyboard with start and end states. For generated shots the two extra columns are required. Each shot has to know where the previous one left things. A simple storyboard to connect every shot covers why and how.
4. Locked assets. One reference image per character (front, side, in costume), one per location, one per important prop. Generate these first, approve them, and then do not touch them. Every later shot references them. Casting, in this workflow, is asset locking.

A useful test before you move on: could someone else generate this film from your documents alone? If the answer is "they would have to ask me things", the documents are not finished.

Generation: one shot at a time, and do not touch the prompt first

The unit of generation is the shot: five to ten seconds per generation, one row of the shot list at a time, with the start and end states written into the prompt and the locked assets attached as references.

Three habits decide whether you end up with a film or a folder of clips:

Change nothing first. When a take is wrong, generate again with the same prompt before you edit a single word. Video models are stochastic; a fair share of "the prompt is wrong" is actually "this take is wrong", and the next take with identical input may be fine. Only after two or three identical takes fail the same way do you know the prompt is the problem.

Then change one line. When you do edit, edit one thing: the action, or the camera, or the lighting. Change three things and you will never know which one fixed it, which means you cannot reuse the fix in the next shot.

Keep a take log. Shot number, take number, what changed, result, accepted or not. It feels like bureaucracy for the first ten shots and becomes the most valuable document you own by shot thirty, because it tells you which kinds of shots this model cannot do, so you can rewrite them instead of rerolling them forever.

How many takes per shot is normal? Hell Grind is the only large project with published numbers: 115,446 generations for the whole film, 88.8 percent of them video, against roughly 570 finished shots, which works out to about 180 generated videos per shot that made the cut. That is a studio pushing for feature quality with fifteen people. Your number for a three-minute film will be far lower, but the shape is the same: the accepted take is a small fraction of the takes, and the fraction is the real cost of the film. That is what the cost section below counts.

Will the character still be the same person in the next scene?

Consistency across scenes is the problem that most often kills short films, and it is mostly solved before generation, in the asset lock. Three habits on top of that:

• Reference every shot, even the easy ones. The shot where the character is a silhouette in a wide still gets the character reference attached. Drift starts in the shots where you thought it did not matter.
• Carry end frames forward. The last frame of shot 7 is a reference for shot 8 when the two are continuous. This is the cheapest continuity tool there is.
• Budget for drift. Some shots will need more takes purely to hold the face or the costume. Mark those shots as high risk in the storyboard before you start (crowds, costume changes, a character seen from a new angle) so the extra takes are already in the plan.

When a shot will not hold, the fix is often upstream: a cleaner reference image, or a shot rewritten so the risky element is smaller in frame. Rerolling the same difficult shot for the fortieth time is the expensive answer.

Audio, editing, export: a folder of clips is not a film yet

The edit is where the film appears. Until then you have a folder of clips that each look fine.

• Dialogue. Decide early whether characters speak on camera or the film is carried by voice-over and sound. Generated on-camera speech has to hold lip sync and voice identity across every line; voice-over only has to hold the voice. For a first film, voice-over plus expressive silent performance is the more forgiving choice.
• Music and sound design. A consistent sound floor (room tone, ambience) is what makes separately generated shots feel like one place. Add it before you judge whether the shots cut together; half of "these two shots do not match" is silence.
• The cut. Assemble in shot-list order, then cut for rhythm. Expect to drop shots. Forty generated shots often become thirty-two in the film, and the take log will tell you which eight were the ones you fought hardest for. That is normal.
• Export. One master at the highest quality you generated, then platform versions. If you need a vertical cut, go back to the storyboard and make one; cropping a landscape master will not give you one.

Will a three-minute short film made with AI cost more than $300?

The formula is short:

total ≈ (number of shots × takes per shot × cost per take) + fixed costs (voices, music, subscriptions, your time)

Everything people argue about is inside "takes per shot". Cost per take is on the model's price list; takes per shot is on your take log.

So start with the price list. We will use two video models that publish a per-second list price: Seedance 2.0 on BytePlus ModelArk, and MiniMax's H3 on the MiniMax open platform, both pay-as-you-go, both billed on seconds of output (official price pages checked 2026-08-25; list prices, no video input, limited-time discounts excluded):

Model	Resolution	Per second	One 6-second take
MiniMax H3	768P	$0.08	$0.48
MiniMax H3	2K	$0.13	$0.78
Seedance 2.0	480p	$0.07	$0.42
Seedance 2.0	720p	$0.15	$0.90
Seedance 2.0	1080p	$0.37	$2.22
Seedance 2.0	4K	$0.78	$4.68

Some details from the price pages matter for the estimate. Seedance bills by token (seconds × width × height × frame rate ÷ 1024), so the same resolution at a different aspect ratio costs slightly more or less; the figures above are 16:9, and only successful generations are charged. H3 takes the first five reference images free and charges $0.04 for each one after that; audio input is free. Seedance 2.5 is listed on the same page at roughly 1.5 times the 2.0 rate.

Now the worked example: a three-minute film of forty shots. Each take is generated at 6 seconds and trimmed to about 4.5 in the edit, so one pass over the whole film generates 240 seconds to end up with 180. You pay for generated seconds; estimating from finished seconds is the most common mistake.

Scenario	Takes per shot	Seconds generated	H3 768P	H3 2K	Seedance 2.0 720p	Seedance 2.0 1080p
Everything works first time	1	240	$19	$31	$36	$89
Careful hobbyist	5	1,200	$96	$156	$180	$444
Demanding hobbyist, several difficult scenes	20	4,800	$380	$620	$720	$1,780
Hell Grind ratio	180	43,200	$3,500	$5,600	$6,500	$16,000

Back to the question in the heading: will a three-minute film cost more than $300? It stays under if you keep to five takes per shot and off 1080p. At twenty takes, even the cheapest tier is past $380.

The first row is the number the marketing implies. The last row is the take ratio a studio needed to reach feature quality, scaled down to three minutes. A first short film usually lands between rows two and three, and the pre-production section above is what moves it toward row two; a better model does little. Every hour spent on the shot list and the asset lock is a row of takes you do not pay for.

One saving can be read straight off the price list. MiniMax prices the takes and the keeper separately: generate at 768P for $0.08 a second, then regenerate only the accepted take into 2K for $0.05 a second (input materials from the original task are billed again; a text-only task, or one with five reference images or fewer, has nothing extra to pay). The takes you throw away are made at the cheap tier; only the one you keep pays for resolution. Seedance has no equivalent upgrade item, but the principle carries: draft at 480p, and generate the final at 720p or 1080p once the shot is settled. That final pass is a new generation, and the frame will differ.

Two costs the formula hides:

• Time. At twenty takes per shot, forty shots is eight hundred generations to review. Reviewing is work. Plan for days.
• Credits that expire. Subscription credits on a monthly clock push you to generate before you are ready. Do the pre-production on paper first; paper does not expire.

Has anyone made a film with AI, and where do you watch them?

Has anyone made a feature film with AI? Yes. Hell Grind (2026) is a 95-minute feature with images generated with AI, made by fifteen people in a fourteen-day generation window for under $500,000, screened on the fringe of Cannes in May 2026, with the whole production process published. There are many shorter ones on the video platforms and in showcases dedicated to this kind of work.

Where can I watch short films made with AI? The model makers' own channels publish workflow films alongside the finished pieces, which is useful because you see the process; curated galleries and festivals collect the rest. Watch a few with the take log in mind: the good ones are usually the ones where the pre-production shows.

---

A short film made this way is a small production, with the same stages a film has always had, and one thing has moved: the money and the effort have left the set and gone to two places, the documents you write before you generate, and the takes you throw away.

Write the documents. Count the takes. Everything else in the workflow is there to make those two numbers smaller.