A year ago, making a short video with open source AI meant stitching together five research repos and hoping the versions matched. Today several projects bundle the whole job, and some run on a laptop. The catch is that each one has a very different idea of who it's for.
This list groups twelve projects by what you are trying to make. The first few are engines for people who want full control. The middle covers voice and music, and the last group is finished apps. For each one I've noted what you need to run it, because hardware and licensing are where most people get stuck.
One rule applies to the whole list. These tools can clone voices and swap faces, so only use a voice or likeness you have permission to use, and label generated media where a platform asks you to.
Every entry below was found through the @RoundtableSpace feed on X, then checked against the project's own page on 4 November 2026. Numbers quoted are the projects' own claims. Read each license and README before you depend on anything here.
| Project | Best for | License |
|---|---|---|
| ComfyUI | Node-based control over image and video pipelines | GPL 3.0 |
| AnimateDiff | Turning image models into animation | Apache 2.0 |
| HyperFrames | Rendering video from HTML | Apache 2.0 |
| Openjutsu | Swapping characters in existing video | MIT |
| OmniVoice | Multilingual text to speech | Apache 2.0 |
| VoiceStudio | A full voice app on your own device | AGPL 3.0 |
| AudioCraft | Generating music and sound effects | MIT (code) |
| MoneyPrinterTurbo | Automated short videos from a topic | MIT |
| OpenCreator | An AI workspace for creators | Apache 2.0 |
| ArtCraft | An IDE for AI image and video work | See repo |
| PhotoCraft | A Photoshop-style editor in Rust | See repo |
| yt-dlp | Downloading your source footage | Unlicense |
1. ComfyUI
ComfyUI bills itself as the most powerful and modular AI engine for content creation. You build a pipeline as a graph of nodes, so every step of a generation is visible and repeatable, and you can save a workflow and share it.
It has a real learning curve, and it rewards the effort. Many other tools on this list sit on top of it, so time spent here carries over. Start with one of the bundled templates, change one node at a time, and only then build your own. There is a comfy-cli package for installation, and the README covers the different GPU setups. A dedicated GPU makes a large difference for video.
2. AnimateDiff
AnimateDiff is the official implementation of an ICLR 2024 Spotlight paper. It is a plug-and-play module that turns most community text-to-image models into animation generators, with no extra training.
That's why it has lasted. You keep the model and style you already like, and motion comes from the module. It's a research repo, so expect to read code and manage dependencies yourself with pip install -r requirements.txt. Most people will meet it through a ComfyUI workflow, which is the easier door. Newer video models now beat it on realism, but it's still worth understanding for stylized animation.
3. HyperFrames
HyperFrames, from HeyGen, turns HTML, CSS, media and seekable animations into deterministic MP4 videos. You can use it locally through a CLI, from coding agents with skills, or as the rendering core behind a hosted tool, and desktop apps are available for macOS, Windows and Linux.
Deterministic is the word to notice. The same source produces the same video, which suits explainers, data videos and anything you'll regenerate with new numbers. Because the source is HTML, a coding agent can write and edit a video the same way it edits a web page. Add it to an agent with npx skills add heygen-com/hyperframes. This one is for motion graphics, not generating realistic footage.
4. Openjutsu
Openjutsu is an open source tool for character and object swaps in video. You drop in a video and a character photo, press Generate, and the main person is replaced by your character. You can also change hair or clothes, remove something, or move the subject elsewhere, while the camera, cuts, motion and soundtrack stay.
It runs on open weights on your own GPU: MiniMax H3 for the video, SAM 3.1 to track people, and Qwen3-VL 8B to describe pictures and shots. The app is a small Next.js front end over a ComfyUI workflow. Only use footage and likenesses you have rights to. Putting a real person into a video without consent can cause real harm and may be illegal.
5. OmniVoice
OmniVoice is a zero-shot text-to-speech model that supports over 600 languages, with voice cloning and voice design. Its authors describe a diffusion language model-style architecture and emphasize inference speed.
The language count is the reason to look at it. Most TTS systems handle English well and everything else poorly, so this is worth testing if you publish in Indonesian or any smaller language. Test your real scripts, because numbers, names and code terms expose weaknesses quickly. It has a Python API and command line tools, and you install PyTorch for your hardware first, as the README describes.
6. VoiceStudio
VoiceStudio is an AI voice studio that runs on your device. It lists cloning, design, dubbing, dictation, transcription and narration in 646 languages, on laptops, desktops, servers, phones and tablets. A Homebrew tap and download page make it far easier to try than a raw model, and paid Pro and Cloud tiers sit alongside the open version.
Pick this if you want an app and not a Python environment. The AGPL license matters if you plan to build it into a service. Local processing is the main appeal for private recordings, so check the settings to confirm what stays on your device.
7. AudioCraft
AudioCraft is a PyTorch library for deep learning research on audio generation, with inference and training code for AudioGen and MusicGen. The README asks for Python 3.9 and PyTorch 2.1.0, so use a dedicated environment.
It's a research library and an older one, so expect some setup friction. It's still a clean place to learn how text-to-music models work and to generate background beds and sound effects. Check the separate license terms on model weights before you use output commercially, because code and weights are often licensed differently.
8. MoneyPrinterTurbo
Give MoneyPrinterTurbo a topic or keyword, and it writes a script, matches stock footage, generates subtitles and background music, and assembles a short video. It offers a web UI and an API, and you can run it with Docker using docker compose -f docker-compose.release.yml up.
Ignore the name. It will not print money. A fully automated channel of generic stock-footage clips is the kind of content platforms increasingly limit, and viewers skip it. Use it as a fast first draft, then replace the script with something only you could write. As a way to study how an end-to-end video pipeline fits together, it's excellent.
9. OpenCreator
OpenCreator, formerly KrillinAI, is an open source workspace that brings visual settings, reusable skills and agents together for scripts, video, images, voice, translation and editing. The README is available in many languages, including Bahasa Indonesia.
Its appeal is having the steps in one place. Translation and dubbing were KrillinAI's original strength, so try it first if you want to repurpose videos for another language. Building from source uses pnpm, and the docs cover provider setup.
10. ArtCraft
ArtCraft calls itself the IDE for artists. You compose in 2D, stage scenes in 3D, and choose the models that fit your work. The goal is to turn prompting into crafting, using visual tools for precise, repeatable results.
If you've ever fought a prompt to move one object slightly to the left, you'll see the point. Staging a scene directly gives you control that text can't. Downloads and GitHub releases are available. I couldn't find a license file at the usual path in this repo, so check the terms before commercial use.
11. PhotoCraft
From the same team, PhotoCraft is an open source, clean-room reimplementation of Adobe Photoshop's feature set in pure Rust. It has layers, masks, adjustment layers, layer styles, type, vectors, brushes and real PSD files in a native offline app.
It's on this list because generated images nearly always need retouching, and a native editor that opens PSDs fits that step. Build and run with cargo run --release -p photocraft -- image.psd. It's young compared with what it replaces, so test your own files. As with ArtCraft, check the licensing in the repo.
12. yt-dlp
yt-dlp is a feature-rich command line audio and video downloader that supports thousands of sites. It's a fork of youtube-dl based on the now inactive youtube-dlc, and nearly every video pipeline ends up relying on something like it.
Editors use it to archive their own uploads, pull reference clips and get source audio for transcription. Download only what you have the right to use, and follow each site's terms. A basic command is yt-dlp URL, and the README documents hundreds of options for format selection and subtitles.
Which ones to start with
Your hardware decides much of this. On a laptop with no GPU, start with HyperFrames, because it renders HTML rather than running a video model, or use VoiceStudio for audio. With a recent GPU, ComfyUI opens up most of the image and video work here. For a phone-first workflow, voice apps are the realistic option today.
Then match the tool to the job:
- Explainers and data videos: HyperFrames.
- Stylized animation: ComfyUI with AnimateDiff.
- Editing footage you already have: Openjutsu.
- Narration in many languages: OmniVoice or VoiceStudio.
- Dubbing an existing video: OpenCreator.
- Rough drafts from a topic: MoneyPrinterTurbo.
Check these before you publish anything
- Separate the license of the code from the license of the model weights. They are often different, and weights may limit commercial use.
- Check where stock footage, music and fonts came from. A tool that fetches them doesn't clear them for you.
- Follow platform disclosure rules for realistic AI media.
- Keep your source files and prompts, so you can show how a piece was made if someone asks.
- Listen to the whole thing. Generated narration can mispronounce a brand name in the one line that matters most.
What these tools still leave to you
None of them decides what is worth saying. The videos that hold attention usually come from a specific idea and a clear point of view, and automation only speeds up the part after that. Use these tools to cut the hours of production, then spend some of that saved time on the script and the first three seconds.
If you generate images often and want prompts to start from, The Prompt Lab: Complete Bundle in my store collects creative and product marketing image prompts.
A simple pipeline for one short video
Here is a workflow that combines a few of these tools. It assumes you have a topic and about two hours.
- Write the script yourself in plain language. Aim for one idea, and keep it short enough to say in under a minute.
- Read it aloud and cut anything you stumble over. Spoken scripts need shorter sentences than written ones.
- Generate narration with OmniVoice or VoiceStudio, or record your own voice if you're comfortable doing so. Your real voice often performs better for personal topics.
- Build visuals. Use HyperFrames for text, charts and diagrams, and ComfyUI for any generated scene.
- Retouch stills in PhotoCraft or your usual editor.
- Add captions, since many people watch without sound.
- Export, then watch the result on a phone before you upload it.
Save the whole project folder. When a video does well, you'll want to remake it with a new topic, and a saved HyperFrames source or ComfyUI workflow makes that quick.
Hardware and cost notes
Running video models locally is the expensive part. They need a capable GPU, plenty of disk space for model weights, and patience. If you don't have that hardware, renting a cloud GPU for an afternoon often costs less than buying one, and it lets you test a workflow before you commit. Voice tools are far lighter, and many can run on an ordinary laptop.
Plan for disk space from the start. Model files are large, and a few checkpoints can fill a small SSD. Keep them on one drive and point every tool at the same folder when the tool allows it, so you don't download the same weights three times.
Also expect things to break. These projects move quickly, and a guide from six months ago may not match the current version. When something fails, the repo's issues page usually has the answer sooner than a search engine. Pin the versions that work for you, and upgrade on purpose when you have time to fix what breaks.
Comments