# Skill: Generate short video clips in ComfyUI using image-to-video workflows

> Lets an AI agent set realistic expectations, frame a sensible task, and guide a user through the image-to-video approach in ComfyUI -- whenever someone wants to animate a still image or create a short motion clip without prior video generation experience.

Source: sodigi·learn -- comfyui/video-animaciya · https://sodigi.io/learn/comfyui/video-animaciya

## When to use
- User wants to animate an existing ComfyUI image (e.g., make eyes glow or hair sway).
- User asks "how do I make video in ComfyUI."
- User is confused about why video generation takes so much longer than image generation.
- User wants to understand AnimateDiff or modern video models without getting lost in model name churn.

## Core rules
- Video is a stack of frames displayed in sequence; the entire challenge is keeping neighboring frames consistent so motion is smooth, not a flickering mess.
- Prefer image-to-video (animate an existing image) over text-to-video for first attempts -- it is more predictable because the first frame is already known.
- Generate short clips (a few seconds) first; extend length by stretching (starting a new clip from the last frame) rather than requesting long clips upfront.
- Video is resource-heavy: expect generation time in minutes, not seconds, even for a two-second clip; this is normal.
- AnimateDiff adds a "motion module" on top of a regular image model to link frames across time; modern video models do this natively -- the tool names change, the principle (link frames in time) stays constant.
- Frame interpolation (filling in-between frames) smooths a jittery raw clip; apply it after generation, not before.
- Describe motion for one or two elements only; asking everything to move simultaneously creates chaos.
- Hardware requirements are significantly higher than for image generation; if VRAM is limited, use a cloud service for video work.

## Procedure
1. Start from a finished still image rather than generating video from text on the first attempt.
2. Choose an image-to-video workflow (AnimateDiff-based or a native video model) compatible with your GPU.
3. Set a short frame count (enough for 2-3 seconds at the target frame rate).
4. Use a small resolution for the first draft; only increase resolution once the motion looks correct.
5. Write motion guidance that specifies only what should move; keep everything else static.
6. Run the generation. Budget minutes of wait time -- this is expected.
7. If the result is jittery, apply frame interpolation to smooth it.
8. To create a longer clip, take the last frame of the finished clip and use it as the starting image for a second clip, then concatenate.

## Ready-to-use prompt
```
I want to create a short animated clip in ComfyUI from an existing image. Help me:

1. Choose image-to-video over text-to-video and explain why it is more predictable for a first attempt.
2. Set a sensible frame count for a 2-3 second clip.
3. Describe only what should move in the scene: [describe the motion here].
4. Know what hardware or cloud resources I will need.
5. Smooth the result with frame interpolation after generation.
6. Extend the clip to a longer duration if needed.

My source image shows: [describe the image].
My GPU VRAM: [X] GB.
```

## Pitfalls
- Asking for a one-minute clip on the first try -- memory runs out and wait time becomes impractical; start with a few seconds.
- Chasing maximum resolution on the first draft -- get clean motion at small size first, then upscale.
- Expecting video generation to finish in seconds like images -- minutes are normal; it is not a freeze.
- Trying to animate every element simultaneously -- limit motion to one or two elements for a stable, readable result.
- Memorizing specific model names for video -- the landscape changes monthly; hold onto the principle (frames + temporal linking), not the names.
- Running video generation on hardware that is only just sufficient for images -- video needs a noticeably larger VRAM headroom.