To make an AI video in ComfyUI, load a video-generation workflow and choose a compatible model, then describe the clip before running the workflow. ComfyUI is an open-source tool that runs AI models in a visual graph. Nodes handle steps such as loading a model, reading a prompt, and creating an output. You can run it on your computer or use a cloud GPU.
For a first clip, try the official Wan 2.2 workflow. It supports text-to-video, which creates a clip from a written prompt, and image-to-video, which animates a starting image. The ComfyUI workflow gallery also offers workflows for other video models.
Choose a workflow and model
Start with a workflow built for the model you plan to use. A workflow connects nodes and settings to tell ComfyUI how to generate a video. Models and workflows must match: a workflow that expects Wan 2.2 needs the corresponding Wan files and nodes.
ComfyUI’s Wan 2.2 tutorial shows a native workflow with nodes that load the diffusion model, text encoder, and VAE. The diffusion model creates video frames. The text encoder turns your prompt into information the model can use. The VAE converts the model’s internal representation into visible images and back.
The model’s size determines the hardware you need. ComfyUI’s documentation says the Wan 2.2 5B version works well with 8 GB of video memory (VRAM) when native offloading is enabled. Offloading moves some work between the GPU and system memory. Check the chosen workflow’s model requirements before downloading files or starting a generation.
Generate a clip with Wan 2.2
Use the following sequence for the official Wan 2.2 workflow:
- Open ComfyUI and load the Wan 2.2 workflow.
- In the Load Diffusion Model, Load CLIP, and Load VAE nodes, select the matching model files. CLIP is the text encoder used in this workflow.
- Choose text-to-video or image-to-video. For image-to-video, load the starting image in the workflow’s image input.
- Edit the CLIP Text Encoder prompt. Describe the subject, setting, and movement you want to see. For example: “A red paper kite rises above a grassy hill as clouds move across the sky.”
- Adjust the size and frame length if those settings are available in the workflow. Larger or longer generations require more GPU memory and processing.
- Select Run, or press Ctrl+Enter, and wait for ComfyUI to finish.
Keep the prompt focused on one scene and one main action while you test the workflow. If the result misses the movement you want, revise the action in the prompt and generate again. Change one setting at a time so you can tell what affected the output.
Fix common setup problems
If a workflow fails to load, check whether ComfyUI reports missing nodes. Some downloadable workflows use community-made custom node packs that add functions beyond ComfyUI’s core features. Install only the pack the workflow needs. ComfyUI Manager is the preferred way to install it. Restart ComfyUI, then look for IMPORT FAILED in the startup log. The custom node pack overview describes this installation approach.
If the workflow loads but cannot find a model, confirm that each model-loading node points to the right file and that the file matches the workflow. If the generation runs out of VRAM, shorten the clip, reduce its output size, or switch to a smaller model your hardware supports. ComfyUI’s Wan examples include a 1.3B model version that requires 8 GB of VRAM.
When AnimateDiff makes sense
AnimateDiff adds motion to images generated with Stable Diffusion. It works with text-to-image and image-to-image workflows. For video-to-video, it works with ControlNet, a tool that guides image generation. Its motion modules target Stable Diffusion 1.5; they do not support SDXL or Stable Diffusion 2.x. See the beginner’s AnimateDiff guide for setup details.
For your first attempts, use the ComfyUI workflow made for a video model instead of adapting an image-generation workflow. Choose AnimateDiff when you specifically want to add motion to a compatible Stable Diffusion setup and have enough VRAM for the intended clip.