walid@portfolio:~/lab/local-video-generation$
cd../lab
03ideaSep 2026

Unlimited, because it runs on your machine

Setting up a free local app that runs open video, image and audio models on your own GPU

One app, installed with a click, running open video, image and audio models locally. The trade is explicit: you supply the GPU and the patience — several minutes per clip on a modest card — and in exchange there is no subscription and no generation limit, forever. The hardware check comes first, because this is the one setup where the answer might genuinely be “not on your machine”.

VideoLocalGPUOpen modelsSelf-hostedPinokioInstaller scriptDesktop launcherHugging Face
i
The trade being made

Everything runs on your own hardware, which is the entire point: no subscription, no credits, no per-generation cost, and no cap on how much you make. The cost moves instead to time and electricity. On a modest card a single short clip takes several minutes — so the workflow is to queue several and walk away, not to sit watching a progress bar.

Before anything else

Check your machine first

This is the step people skip and then waste an hour on. Hardware decides whether the rest of the guide applies to you at all.

A Windows or Linux PC with an NVIDIA card, roughly GTX 10-series or newer. Recent AMD cards also work.At least 6 GB of VRAM. Some models run at exactly that; more VRAM buys you longer clips and higher resolution, not just speed.Around 60 GB of free disk space — the model downloads are large, and they accumulate as you try more of them.No discrete GPU, or on a phone? This route is not for you — use a hosted option instead.
Step 1

Install the one-click installer

Pinokio exists to install awkward AI tooling without a terminal. It is the layer that makes the rest of this a click rather than a dependency fight.

1

Download it

Pick the Windows or Linux build and install it like any normal application.

2

Get past the safety warning

Windows will likely flag an unrecognised publisher. Choose to run it anyway — expected for a small independent tool, though it is your call to make knowingly.

3

Point it at your roomiest drive

First-time setup asks where to store files. Choose the drive with the most space free: this is where tens of gigabytes of models will land.

Step 2

Install the app inside it

1

Open the discover page and search for it

The app browser lists community install scripts alongside official ones.

2

Pick the right script, not the default listing

The maintained community installer is the one to use — the project’s own documentation recommends it over the default entry. See the note below about which variant you need.

3

Download, then install, then wait

It sets up everything automatically and takes a while. Let it finish rather than intervening.

4

Start it and open the guides tab first

The app opens in your browser. Its built-in guides recommend which models suit your specific GPU — worth two minutes before you download a model that will not fit.

!
One thing worth checking yourself

The guide sends AMD users to a separate installer variant and tells Mac users this is not for them yet. The community installer it recommends, however, describes itself as a unified installer covering NVIDIA, AMD and Apple silicon, and names specific supported AMD cards. That suggests the variants may have been merged since the guide was written. Check the installer’s own description before hunting for a separate AMD script — and if you are on a Mac, it may be worth a look rather than an automatic no.

Step 3

Pick the model for the job

All of these live in one dropdown. The first time you select a model it downloads automatically — a one-time wait per model, not per generation.

What you wantWhat to select
An image turned into a cinematic videoThe general image-to-video model — or the audio-capable one if you want sound generated alongside it.
Video with sound and music in one passThe audio-capable video model. Choose its pruned smaller variant if VRAM is tight.
The same character across many shotsThe reference-image variant, with your reference images supplied.
Your own footage animated, or a character copying your movementThe animation models — give one a character image plus your driving video.
Anime and stylised scenesThe general video model plus a style LoRA, browsable from the built-in model browser.
Low-quality footage made sharpAdvanced mode → post-processing → one of the upscaling models.
Fast, lightweight videoThe lighter-weight video model.
Still imagesOne of the two image models.
Music, or sound for an existing clipThe music model, or the audio-to-video tool that scores footage you already have.
Step 4

Your first generation

1

Select a video model and leave the defaults alone

Around 480p and five seconds. Resist raising either until you know how long your card takes at the baseline.

2

Add an image if you have one, then describe the motion

Prompts here are about camera and atmosphere rather than subject — a slow push-in, the quality of the light, what moves in the scene, what it sounds like.

3

Generate, then go and do something else

Several minutes per clip on a 6-8 GB card is normal, not a fault. There is a built-in queue: stack up several and come back.

!
It is free — so anyone charging you for it is not selling you it

The software is genuinely free and open source. Sites that charge for access to it, or for an “installer”, are not affiliated with it. Stick to the project’s own repository and its official site, and get the installer through the app browser rather than from a search result. The project’s community chat is the place to go when something breaks — it is active, and it is free too.

No GPU?

The fallbacks

Several of these models have official hosted apps you can use in a browser, with the usual free-tier limits.The model weights themselves are published openly, so renting a cloud GPU later is a real option once you know which model you actually want.Neither route gives you the unlimited-generation economics that make the local install worth the setup — that only comes from owning the hardware.
i
Provenance

The hardware requirements, install sequence and model-selection guidance follow the source guide. Of the projects it points to, the community installer script and the alternative desktop launcher were both confirmed on GitHub on 16 September 2026, as was the one-click installer’s own site. The project’s canonical repository did not surface through the search index available here, though an ecosystem of forks and companion projects around it clearly exists — so treat the repository link as the guide gives it and verify it yourself before installing. Specific model names and version numbers are deliberately described by role rather than pinned, since they turn over quickly and the built-in guides recommend the right one for your card anyway.

← previous
Rebuild it, or keep paying
next →
Six analysts, one report
← all experiments