Local video generation
Open text-to-video has recently become runnable on consumer hardware. What the models do, which ones fit your GPU, and the trade-offs in memory, length, and licensing.
What you’ll learn
How local video generation works, which open models you can realistically run on your hardware, and where the practical limits are.
The state of local video
Until recently, open text-to-video needed data-centre hardware. That has started to change: there are now models that generate short clips on a single consumer card. It is still the most demanding thing you can do locally, so expectations matter, but it is no longer out of reach.
How it works
Video generation extends diffusion, the same technique behind local image models, from a single image to a sequence of frames that move coherently. That coherence is what makes it expensive: a video model is much larger and far more memory- and compute-hungry than an image model. Crucially, its hardware needs are driven by resolution, the number of frames, and the number of steps, not by anything like a language model’s context. Most models generate text-to-video from a prompt, and many also do image-to-video, animating a still image.
What you can run
- Wan 2.2 TI2V-5B is the one to try first. At five billion parameters it fits a single 24GB card with the offloading flags, does both text-to-video and image-to-video at 720p, and is Apache 2.0, so there are no commercial conditions.
- LTX-Video is built for speed, with a small 2B variant aimed at modest VRAM. Its weights use a custom, version-specific licence, so read the terms for your checkpoint before commercial use.
- HunyuanVideo produces higher quality but needs 45GB or more, and ships under a conditional community licence. It is a workstation or rented server model, not a laptop one.
How to run it
The most common route is ComfyUI, which has node-based workflows for each of these models, exposing the sampler, resolution, and frame count directly. The models can also be run from their own repositories or through Hugging Face diffusers.
What can go wrong
- Running out of memory. VRAM use rises sharply with resolution and frame count. If a model will not fit, lower the resolution or the number of frames before anything else.
- Expecting long clips. Open models generate a few seconds at a time. Longer video means stitching clips, not a single generation.
- Slow generation. Even when a model fits, a clip can take minutes. The distilled and smaller variants trade quality for speed.
Next steps
Video generation shares its tooling and diffusion foundation with local image generation, which is the gentler place to start. For the audio side of generative media, see local music and audio generation.