A local-first AI filmmaking studio you drive from your coding agent. Short clips chained and cut into pieces that run for minutes, rendered on a swarm of consumer-grade GPUs — which is why a ten-second shot costs cents rather than dollars.
Hosted, or free and self-hosted with your own API key — same studio either way
Run it yourself and nothing leaves your disk: the project is a directory you own, the assets are files you can open, and compute is billed to your own key at cost. Or use the hosted version and skip the setup — pay-as-you-go, no subscription, and the same studio underneath.
An embedded terminal runs your coding agent inside the project. You describe the film; it plans shots, composes references, and renders clips through the studio's tools.
A production packet holds characters, props, locations, and style with stable IDs. Every clip carries a contract for what must be true at its start and what it promises the next cut.
Image, video, voice, and upscale all run on an open-model fleet — FLUX, LTX, Qwen — priced per task. Every model is selectable; you are not locked to one vendor's idea of the right one.
Node 20+, ffmpeg, and a coding agent you already use.
git clone https://github.com/conrader/filmhutnode scripts/deapi-doctor.mjs confirms the key works and
lists the models your account can see. It costs nothing../scripts/start.sh, then open the canvas and tell the agent
what you want to make.# the whole setup
git clone https://github.com/conrader/filmhut && cd filmhut
cp .env.example .env # then paste your key, in quotes
node scripts/deapi-doctor.mjs # free preflight
./scripts/start.sh # open the studioVideo generation is priced by whoever owns the silicon. filmhut doesn't rent time on hyperscale accelerators — it harnesses thousands of idle consumer-grade GPUs, the cards already sitting in gaming machines around the world, and pays for the work rather than the hardware.
That is the whole reason a ten-second clip costs cents instead of dollars. The same open models — FLUX, LTX, Qwen — running on distributed consumer hardware, priced per task, with no idle capacity to amortise and no reserved-instance minimum to cover.
No model generates a whole film. Every one of them has a hard ceiling — ten seconds on the model filmhut defaults to, four on the cheaper one — and that ceiling is a frame budget, not a time budget, so asking for more just clamps. Length comes from assembly: render shots, order them, cut them, stitch them.
Each shot opens on the previous one's last frame, so motion and framing carry across the join. The timeline holds the order, trims each clip to the part that actually works, and exports a single MP4 — stream-copied losslessly where the cuts fall on keyframes, re-encoded exactly where they don't. Three minutes is eighteen shots. An hour is around three hundred and sixty.
A ten-second generation that is good for six gets cut, not regenerated. That is the difference between paying once for a shot and paying four times for it.
Characters, props and locations get stable IDs in a production packet, and every clip carries a contract for what must be true at its start and end. Drift becomes something you can catch at the shot rather than at the edit.
Stitching is ffmpeg on your own machine. No upload, no render queue, and the master never leaves your disk.
What actually degrades over length. Chaining carries composition well and identity poorly — references are frames, not trained identities. In our own 30-second test the stronger model held a recognisable character across three shots; the cheaper one had lost her by the third. Expect to rebuild a character's opening frame from a reference sheet every few shots rather than chaining indefinitely, and to regenerate the occasional shot. Longer pieces are a matter of more shots and more inspection, not a longer render.
Prices read from the live pricing endpoint, not a rate card. You pay the compute supplier directly; filmhut takes nothing.
| Piece | Built from | Compute |
|---|---|---|
| One six-second establishing shot | a frame, then a clip | $0.05 |
| 30 seconds, four shots | four frames, four clips | $0.21 |
| Three minutes with dialogue and score | 18 shots, lip-synced, music | $2.40 |
One number those figures leave out. Every shot also costs a turn or two of your coding agent — planning, composing references, reviewing what came back. At feature length that can exceed the GPU cost. It comes out of whatever agent subscription you already have rather than a bill from us, but it is real, and you should know about it before you plan an hour of footage.
Worth knowing before you clone it.
Around ten seconds on the default model — a frame budget, not a time budget. Longer scenes are chained, using one clip's last frame as the next one's first.
References are frames, not trained identities. Holding a character across many shots means building a sheet and composing from it — the continuity tools help, but drift is real.
You run it. There is no account, no subscription, and no upload. That is the trade: more setup, in exchange for owning the result.
Same tool. The difference is who holds the machine and the key.
Sign up and start. Credit is pay-as-you-go, reserved before a job runs and settled at the real price; a failed generation costs nothing. No subscription.
Clone it, bring your own key, and nothing leaves your disk. Free, and the version where your project files are genuinely yours.
Free, source-available, and yours to run. Built on pai-pro by Utopai Studios.
