Skip to content
Agents tracked: 492 Releases (7d): 545 Agent status: 1 with issuesSince last limit reset: Claude 9d · Codex 7d For AI agents: Markdown Updated Nov 20, 2026

✓ Checked

AI Video Generation: 10 Open-Source GitHub Repos for Filmmakers

A viral list of 10 open-source repos for AI filmmakers, checked on GitHub: the six new to us, which need a serious GPU, how model licenses differ from code licenses, which projects have gone quiet and where to start.

On October 26, 2026 Rimsha Bhardwaj (@heyrimsha) posted a second list of 10 open-source GitHub repos, this time for "AI filmmakers": tools that generate video, animate photos, sync lips and automate editing.

Her October 16 list is already on AgentGid as AI Video Editing: 10 Open-Source GitHub Repos, Checked, and three repos come back from it: OpenChatCut, FireRed-OpenStoryline and AutoSubs. LTX-2 is in our AI video catalog. So this check is mostly about the six that are new to us: ComfyUI WanVideo Wrapper, CogVideoX, VACE, DiffSynth-Studio, MuseTalk and Editly. We read each repository's README and license and pulled the numbers from GitHub's API on October 27, 2026.

Four kinds of project on one list

The post presents all ten as one kind of thing. They are not: four are models you download, two are toolkits that run other people's models, two are editing agents, and two are helpers, one of which has no AI in it at all.

Models you downloadLTX-2 (video with sound), CogVideoX (text or image to video), VACE (create and edit with references and masks), MuseTalk (lip sync).
Toolkits that run modelsComfyUI WanVideo Wrapper (ComfyUI nodes for Wan and related models) and DiffSynth-Studio (inference and training for many image, video and audio models).
Editing agentsOpenChatCut and FireRed-OpenStoryline: you describe the edit and an agent works the timeline.
HelpersAutoSubs makes subtitles on your machine for your editor. Editly assembles clips, titles and music from a JSON spec with Node.js and ffmpeg; it generates nothing.
Our grouping of the ten repositories, from their own READMEs.

The ten, checked

Repo Stars License Latest activity Note
LTX-2 9,895 LTX-2.x Community License (GitHub shows NOASSERTION) v1.5.0, Oct 18, 2026 In our catalog
ComfyUI WanVideo Wrapper 6,749 Apache-2.0 last push May 24, 2026; no releases 1,285 open issues
CogVideo 13,114 Apache-2.0 code; CogVideoX-5B under its own license v1.0, Nov 8, 2024; last push Nov 4, 2025 In our catalog
VACE 3,996 Apache-2.0 code; each model keeps its base model's license last push Oct 17, 2025; no releases ICCV 2025 paper code
DiffSynth-Studio 13,309 Apache-2.0 last push Oct 26, 2026 The busiest of the six
MuseTalk 6,805 MIT code (GitHub shows NOASSERTION) last push Sep 26, 2025; no releases Tencent Music's Lyra Lab
OpenChatCut 2,694 AGPL-3.0 v0.2.17, Oct 15, 2026 Checked in our editing guide
FireRed-OpenStoryline 3,566 Apache-2.0 last push Jul 31, 2026; no releases Checked in our editing guide
AutoSubs 4,457 MIT v3.11.0, Sep 29, 2026 In our catalog
Editly 5,556 MIT v0.15.0-rc.1, Jan 19, 2025; last push May 12, 2025 No AI; command line and JavaScript API

Stars, licenses and dates from GitHub's API on October 27, 2026; license details from each README and license file.

What the post gets right

The one-line descriptions match the READMEs. LTX-2 is the official package for a model that generates synchronized audio and video, with a LoRA trainer. VACE does reference-to-video, video-to-video and masked editing in one model. MuseTalk changes a face to match new audio, and its README says it works with audio in several languages. Editly does combine clips, images, titles, transitions and music from a command line or an API. All ten repositories are public and carry a license.

What the post leaves out

You need a serious GPU for six of them. The LTX-2 README says the files for its quick start come to roughly 66 GiB. CogVideoX's own table gives about 1000 seconds on a single A100 for a 5-second clip from its newest model, as reported by the team; its smallest model is listed from 4GB of GPU memory with optimizations on. MuseTalk reports real-time output at 30fps+ on an NVIDIA Tesla V100.

Code license and model license are not the same thing.

  • LTX-2.x weights come under a community license: free to use, but entities with annual revenues of at least $10,000,000 must get a paid license. The license claims no rights in the videos you generate.
  • In the CogVideo repo the code and the CogVideoX-2B model are Apache 2.0, while CogVideoX-5B, including its image-to-video model, is under a separate CogVideoX license. Read that file before commercial use.
  • VACE says each of its models inherits the license of the model it was built on. The Wan2.1-based ones are Apache 2.0; the one built on LTX-Video 0.9 carries a RAIL-M license with use restrictions.
  • MuseTalk's code is MIT and its README says the trained model may be used for any purpose, even commercially, but its test data is for non-commercial research only and the models it depends on (Whisper, a VAE, face and pose detectors) keep their own licenses.

Several are quiet. CogVideo's last release is from November 2024, and its README now points to the team's commercial QingYing service for larger models. VACE and MuseTalk have not been pushed to in a year. Editly's newest release is a release candidate from January 2025.

The WanVideo Wrapper is a personal sandbox by its author's own words. Its README says that if a Wan model already works in native ComfyUI, you should use that instead, and that new accounts are blocked from opening issues because too many people mistook the repo for a video generation service. It has 1,285 open issues.

CogVideoX only understands English prompts. Its README says so and suggests translating other languages with a language model first.

Stars are not activity

DiffSynth-Studio
13,309
CogVideo
13,114
LTX-2
9,895
MuseTalk
6,805
WanVideo Wrapper
6,749
Editly
5,556
AutoSubs
4,457
VACE
3,996
OpenStoryline
3,566
OpenChatCut
2,694
GitHub stars on October 27, 2026. Highlighted bars: no push to the repository since November 2025 or earlier.

Of the six repos new to us, only DiffSynth-Studio was pushed to in the last month, and LTX-2 shipped v1.5.0 on October 18. DiffSynth-Studio added LTX-2.5 support on September 10, 2026, with low-VRAM inference, and its README says its memory management lets consumer GPUs run large models. It also asks for patience: two developers do most of the work, so new features and answers to issues come slowly, as the README itself puts it.

If you work with a coding agent

DiffSynth-Studio is the one on this list that publishes Agent Skills for AI, a set of skills for integrating new models. OpenChatCut has agent skills and an MCP server inside the editor; see our editing guide. Editly's JSON5 edit spec is plain enough for an agent to write and check, which makes it a reasonable last step after a generation model, though its latest release is still a release candidate from January 2025.

For open video models with their licenses side by side, see AgentGid's AI video catalog, which also lists Wan 2.2 and HunyuanVideo 1.5.

Where to start

  • Video with sound from one model: LTX-2, after reading the community license if your company is large.
  • A node workflow for Wan models: native ComfyUI first; the WanVideo Wrapper only for models native support does not have yet.
  • One toolkit for many models, including training LoRAs: DiffSynth-Studio.
  • Editing a shot with references and masks: VACE, with the license of the base model you choose.
  • Lip sync for dubbing: MuseTalk, keeping its dependencies' licenses in mind.
  • Assembling clips from a script: Editly, or Remotion if you want an actively maintained framework.

We did not run these tools; everything above comes from the repositories, their license files and GitHub's API on October 27, 2026.

Was this useful?