What counts as an open source AI video generator
Most tools sold as AI video generators run closed models on someone else's servers. An open source or open-weight video model is different: the developer publishes the trained weights, so anyone can download them and run them on their own hardware. Licenses vary. Some are fully permissive, like Apache 2.0, while others are community licenses with conditions on company size, region or commercial use. In practice most people run these models through ComfyUI or Hugging Face Diffusers on a local NVIDIA GPU or a rented cloud GPU.
Wan 2.2 and Wan 2.1 (Alibaba)
Wan is the most widely used open video family. Wan 2.2, released in July 2025 under Apache 2.0, includes 14B text-to-video and image-to-video models and a smaller TI2V-5B model that makes 720p video at 24 frames per second. The Wan team says the 5B model needs at least 24 GB of VRAM, such as an RTX 4090, and takes under nine minutes for a five-second 720p clip. The older Wan 2.1 has a 1.3B text-to-video model that needs about 8 GB of VRAM and suits 480p. Newer Wan versions are offered through Alibaba's cloud rather than as downloads; the official Wan GitHub organisation publishes weights up to Wan 2.2.
HunyuanVideo 1.5 (Tencent)
HunyuanVideo 1.5 is an 8.3B model released in November 2025 with text-to-video and image-to-video at 480p and 720p, plus an upscaler to 1080p. Clips run about five seconds by default, and Tencent lists a minimum of roughly 14 GB of VRAM with offloading. It uses the Tencent Hunyuan Community License, which does not apply in the EU, the UK or South Korea and requires a separate license for services with more than 100 million monthly users. Read the license before using it commercially.
LTX-2 (Lightricks)
Lightricks' LTX-2 line generates video and synchronised audio together, with text-to-video, image-to-video, keyframes and video-to-video modes, and native output around 1080p with upscaling to 4K. It is released under the LTX-2 Community License, which allows free commercial use for organisations under 10 million dollars in annual revenue; larger companies need a paid agreement. The full download is large, so expect to use reduced-precision weights and offloading on consumer GPUs.
CogVideoX, Mochi 1 and other open models
CogVideoX from Zhipu comes in a 2B version under Apache 2.0 and 5B versions under its own license, making short clips of about six seconds; with memory optimisations it runs on modest GPUs. Genmo's Mochi 1 is a 10B text-to-video model under Apache 2.0 that outputs 480p and runs in under 20 GB in ComfyUI. Others worth knowing include Open-Sora 2.0 (Apache 2.0), StepFun's 30B Step-Video-T2V (MIT, needs an 80 GB GPU), Sand AI's MAGI-1 (Apache 2.0, with a 4.5B version that fits a 24 GB card) and Skywork's SkyReels models, which can extend clips well beyond the usual five seconds.
How FacelessGenie uses open models
FacelessGenie mixes open and closed models and picks per format. Our free tier runs on open models on our own GPUs: Z-Image Turbo from Alibaba's Tongyi lab (Apache 2.0) for scene images and Kokoro (Apache 2.0) for narration, with Whisper timing the word-by-word captions. When a scene is generated straight from text, Wan 2.2 is our default text-to-video model. For higher quality tiers you can switch to closed models such as Veo 3.1, Kling 3 Omni, Seedance 2.0 or Grok Imagine. You get the benefit of open models without installing anything, and the app also writes the script, assembles the scenes, mixes music and renders the final MP4.
Open source text to video vs a finished video
Open-weight models answer one question: what should these few seconds look like? A video someone watches to the end answers many more. What is the story, who narrates it, how do scenes connect, where do captions sit, what music fits? With a local setup you answer each of those with a different tool. That is a good way to learn how AI video works, and a slow way to publish every day. FacelessGenie packages the whole chain into formats such as AI image stories, Reddit-style stories, stickman explainers and talking objects, so the open models do their part inside a complete workflow.
Local AI video generator hardware, in plain numbers
Graphics memory decides what you can run. Around 8 GB handles Wan 2.1's 1.3B model at 480p. Roughly 14 GB with offloading runs HunyuanVideo 1.5. A 24 GB card such as an RTX 4090 runs Wan 2.2's 5B model at 720p, at several minutes per five-second clip. The 14B and larger models are built for 80 GB data-centre GPUs, although community tools squeeze some of them onto smaller cards with quantisation and offloading. If you do not own that hardware, renting a cloud GPU by the hour is the usual middle ground.
Open source AI video and licenses
Open weights do not always mean unrestricted use. Apache 2.0 and MIT models can be used commercially with few conditions. Community licenses add rules: some exclude certain regions, some cap company size or monthly users, and some require derivatives to keep the same license. Image models matter too. FLUX.1 schnell is Apache 2.0, but FLUX.1 dev ships under a non-commercial license for the weights. If your channel is monetised, keep a note of which model and license produced each asset.
Running open models yourself vs FacelessGenie
| FacelessGenie | Local open-source setup |
|---|
| Cost | One free video a day; paid credits shown before each render | Free software; you pay for the GPU or cloud time |
| Setup | None, runs in the browser | ComfyUI or Diffusers, model downloads, drivers |
| Output | A finished MP4 with script, voice, scenes, captions and music | Short clips of about five seconds each |
| Length | From short vertical clips to 20-minute 16:9 videos | Whatever you edit together by hand |
| Models | Open models on the free tier, plus Veo, Kling, Seedance and Grok on paid tiers | Any open model your hardware can hold |
| Control and privacy | Scene-level edits in the app | Full control of every parameter; files never leave your machine |