Video generation
How video generation differs from pictures, the kinds of video graphs, and where to read on.
In plain words
To a video model a clip is a stack of frames drawn all at once, so that the motion holds together. The latent has a time dimension as well as width and height, and the VAE compresses frames in time too. Everything a picture graph has is still there: loaders, text encoder, sampler, VAE decode. What is new: a length, a frame rate, a node that turns frames into a clip, and much larger memory needs.
The kinds of video graphs
| Kind | Starts from | Where to read |
|---|---|---|
| Text to video | A prompt and an empty video latent | Lesson: Video from text |
| Image to video | A picture as the first frame | Lesson: Video from a picture |
| First and last frame | Two pictures, the clip fills the gap | First and last frame |
| Video to video | Another video's motion, depth, pose or camera | Video to video |
| Animation by pose | A character picture and a driving video | Pose animation |
| Talking portraits | A portrait and a voice recording | Talking portraits |
| Video with sound | A prompt; sound is generated too (LTX-2, Ovi) | Audio |
Where you meet it in a graph
- Empty video latents: EmptyHunyuanLatentVideo (Wan 14B, HunyuanVideo), Wan22ImageToVideoLatent (Wan 2.2 5B), EmptyLTXVLatentVideo (LTX).
- Start nodes that attach a picture: WanImageToVideo, WanFirstLastFrameToVideo and others.
- After VAE Decode the frames are an
IMAGEbatch. Create Video joins them with anfps(and optionalaudio) into aVIDEO; Save Video writes the file. Load Video and Get Video Components do the opposite. - The families the app knows for video: Wan 2.1, Wan 2.2, HunyuanVideo, LTX-2.
What changes compared with pictures
- Length and frame rate have rules per model: Video length and frame rate.
- Prompts describe motion and camera: lesson.
- Memory and time grow with width × height × frames: start at 480p and 49 to 81 frames. See Graphics card memory.
- Speed-up LoRAs matter more: Speed-up LoRAs.
Typical mistakes
A picture empty latent in a video graph; fps that does not match the model; a clip too long for memory; t2v models for image to video.
How the app helps
The lesson groups Video, Video: the anchor picture, WanVideoWrapper and Packs and tools build video graphs on Wan, LTX, HunyuanVideo and Kandinsky (see Lessons), and node descriptions cover the core loaders and samplers.