Contents
ComfyUI AI Video Generation Tutorial: Text-to-Video and Image-to-Video
ComfyUI can generate more than images. Recent open source video models, such as the Wan series, LTX-Video, and HunyuanVideo, can run in ComfyUI to generate short videos a few seconds long.
Two main methods
| Method | Input | Best for |
|---|---|---|
| Text-to-video (T2V) | Prompt | Generating a scene from scratch |
| Image-to-video (I2V) | One image + prompt | Making an image “move” with the most control over the scene |
In practice, image-to-video is the most common approach: first use an image model to create a scene you like (or use your own photo or illustration), then generate motion. This usually gives you better quality and consistency.
Hardware requirements
Video generation is very resource-intensive:
- GPU memory: 16 GB or more is recommended. With 12 GB or less, use a quantized or smaller model and lower the resolution and frame count.
- System memory: 32 GB or more is recommended.
- Time: Depending on your hardware and settings, a video a few seconds long may take several minutes to tens of minutes.
Many video models offer different sizes (for example, versions with fewer parameters), as well as FP8 and GGUF quantized versions. Choose these versions first if you are short on memory.
Start with an official template
Video model workflows are more complex than image workflows, so starting with a template is strongly recommended:
- Menu → Browse Templates → Video category.
- Choose a model and method, such as that model’s “image-to-video” template.
- Download the required model files as prompted (diffusion model, text encoder, VAE, and sometimes an image encoder for image-to-video), and place them in the corresponding folders.
- Load a starting image in the Load Image node, edit the prompt, and click Run.
Key parameters
| Parameter | Description |
|---|---|
| Width / height | Video resolution. Start by testing at a lower resolution (for example, around 480p) |
| Frame count (length / frames) | Video length = frame count ÷ frames per second. Some models require a specific frame count (for example, a multiple of 4 plus 1) |
| fps | Frames per second. Use the value recommended by the model (commonly 16–24) |
| steps, CFG | Use the values recommended by the template |
Frame count and resolution have the biggest impact on memory and generation time. Start with short, small settings to confirm the motion looks right, then increase quality.
Prompts
For video prompts, describing motion and camera movement is more important than simply describing the scene:
A girl with an umbrella walks slowly along a rainy street at night, neon signs reflecting
on the wet ground. The camera slowly follows her from behind. Light rain falls.- Describe the subject’s actions (walks, turns her head, smiles).
- Describe camera movement (slow zoom in, pan left, static camera).
- For image-to-video, you don’t need to repeat what’s already in the image. Focus on how it should move.
Save videos
Templates usually output MP4 / WebM / GIF using Save Video, Save WEBM, or VideoHelperSuite’s Video Combine node. Files are saved in the output folder.
Speed-up tips
- Acceleration LoRAs (such as Lightning types) can reduce the step count to a single digit and greatly shorten generation time.
- Use FP8 or GGUF quantized models.
- Generate at a low resolution first, then upscale if you like the result (there are dedicated models and workflows for video upscaling, too).
Set realistic expectations
- Most current open source video models generate clips only a few seconds long. For longer videos, generate clips in segments and edit them together.
- Complex motion, hands, and text are still prone to errors. Generate several versions and choose the best result.
- Models are improving very quickly, and better ones may appear within a few months. Check the templates regularly for new options.
When generating videos of real people, follow the law and platform policies. Don’t create content that impersonates someone or misleads viewers.
Updated October 12, 2026