Local video generation techniques
Video generation on your own hardware is less about picking the right model than about picking the right approach. This section sits under video, and getting started with local video covers the setup this assumes, if you have not generated a clip yet.
Why approach matters more than model choice
Models come and go. The three approaches below have been stable for as long as local video has existed, and a new model almost always slots into one of them rather than replacing them. Once you know which approach fits the shot you want, the model pages tell you which current model is shaped for it, and the pages linked here tell you what to expect from the result.
The approach also decides your costs. It sets how much VRAM the run needs, how long a clip you can hold in memory at once, and how much control you get over what appears in frame. Choosing badly here wastes far more time than choosing the wrong model within an approach.
Text to video
Text to video generates a clip from a prompt and nothing else. It is the most direct approach and the hardest to control: with no starting frame, the model decides composition, subject, lighting and motion all at once, and you only find out what it chose after the run. Good for abstract motion, establishing shots and ideas you cannot draw, frustrating when you have a specific picture in mind.
Image to video
Image to video starts from a still you already have and animates it. Because composition, subject and colour are settled before any motion exists, you are asking the model one question instead of five, and you can iterate on the still cheaply in a still-image workflow before committing to a video run. This is the approach most people end up using for anything that has to look like a particular thing. See image generation for the half of the workflow that happens before the clip.
Frame interpolation
Frame interpolation is not a way of generating content at all. It takes a clip you already have and invents the frames between the existing ones, so the motion reads as smoother or the clip runs longer at the same speed. It is a finishing step, it is the cheapest of the three by a wide margin, and it is usually the first thing to try when a result is nearly right.
How they combine
The common workflow runs all three in order. Generate or choose a still, animate it, then interpolate the short clip that comes out. The video run is the expensive, least controllable step, so you fix as much as you can on either side of it: reroll the still until the subject is right, keep the animation run short, and let interpolation carry the smoothness.: reroll the still until the subject is right, keep the animation run short, and let interpolation carry the smoothness. Text to video enters this chain when you have no still and no way to make one, or when the motion itself is the point.