# Skill: Generate video (text-to-video and image-to-video) with Wan in ComfyUI

> Create short video clips locally - either from a text prompt or by animating an existing still image - using the open-weights Wan 2.1/2.2 models in ComfyUI, with realistic expectations about length, VRAM and render time.

Source: sodigi-learn - comfyui/wan-video - https://sodigi.io/learn/comfyui/wan-video

## When to use

- You need a short video clip generated locally (open weights), not via a paid API.
- You have a still image (often generated with FLUX.2 or Qwen-Image) and want to animate it.
- You want text-to-video for a scene you cannot easily produce as a still first.
- A user expects long, audio, 1080p one-click video - clarify that this is the closed Wan 2.5 API, not local.

## Core rules

- Wan 2.1 / 2.2 (Alibaba) are open-weights and ComfyUI-native; they do text-to-video (T2V) and image-to-video (I2V).
- Prefer image-to-video for control: make a strong still with FLUX/Qwen first, then animate it with Wan.
- Wan 2.2 TI2V-5B = one compact checkpoint for both modes, lighter VRAM; A14B = higher quality but heavier (separate T2V/I2V checkpoints).
- Wan 2.5 (audio, 1080p, longer clips) is a CLOSED API - weights are not public; do not promise local use.
- Set realistic expectations: open local video = short clips (a few seconds), meaningful render time, high VRAM.
- Weak GPU -> use TI2V-5B or a GGUF/quantized build.
- Keep the described motion simple; one clear movement renders cleaner than many simultaneous actions.

## Procedure

1. Update ComfyUI (Wan is natively supported with templates).
2. Download Wan 2.2: TI2V-5B for modest GPUs, A14B for strong GPUs, GGUF for very weak ones.
3. Pick the mode: image-to-video (recommended for beginners) or text-to-video.
4. Open the matching Wan template.
5. For I2V, supply your still image plus a short description of the motion; run and allow render time.

## Ready-to-use prompt

```
Animate this into a short clip with Wan in ComfyUI (image-to-video preferred).
Input still: <attach the generated image>
Desired motion: <one simple, clear movement, e.g. "steam rises and coffee gently pours">
Constraints: short clip (a few seconds), keep motion simple and physically plausible.
If audio / 1080p / long duration is required, note that this needs the closed Wan 2.5 API, not local Wan 2.1/2.2.
```

## Pitfalls

- Running the heavy A14B on a weak GPU (out-of-memory) instead of TI2V-5B or GGUF.
- Expecting long, cinematic, 4K clips - local open video produces short clips and takes time.
- Starting with text-to-video (less control) instead of animating a good still via image-to-video.
- Trying to download Wan 2.5 weights - they are closed; only 2.1/2.2 are open.
- Describing many simultaneous motions, which degrades the result.