Video generation is the most demanding workload you can run locally. It does everything an image model does, for every frame, and has to keep the frames consistent with each other.
What drives the cost
- The model. A video backbone plus a large text encoder and a VAE, all of which must fit or be moved in and out of memory. Some text encoders alone run to several billion parameters.
- Resolution. Pixels per frame, as for images (image guide).
- Frame count. More seconds of video means more frames in memory at once, so memory and time grow quickly with length.
- Steps. More denoising steps give cleaner results and take proportionally longer.
Time, not tokens
Video models are judged in seconds or minutes per clip rather than tokens per second. Our estimates are scaled from a reference run on a card of about 83 TFLOPS of fp16 compute, so a faster card finishes proportionally sooner. Short clips at modest resolution are realistic on a high-end consumer card; longer, higher-resolution clips can take a long time or need heavy offload.
Making it fit
- Pick a smaller or lower-precision variant. Many video models ship fp8 or GGUF versions that keep most of the quality.
- Shorten the clip and lower the resolution first; both cut memory sharply.
- Allow offload. Moving the text encoder or parts of the backbone to RAM works but is slower (why).
- Close everything else that uses the GPU.
Every video model page shows memory and estimated time for your hardware, with a rating in the same terms as everything else (ratings). ComfyUI is the main way to run them locally.