Wan2.2 Google Colab — Complete Video Generation Suite

A comprehensive Google Colab notebook covering every generation mode in Wan2.2 — the open-source MoE video generation model from Alibaba. Includes text-to-video, image-to-video, speech-to-video, pose-driven animation, character replacement, and more.

---

Supported Generation Modes

Section	Mode	Model	Resolution	Notes
2	Text-to-Video	T2V-A14B	480P / 720P	MoE 27B total / 14B active
3	Image-to-Video	I2V-A14B	480P / 720P	Aspect ratio follows input image
4	Text+Image-to-Video	TI2V-5B	720P @ 24fps	Runs on RTX 4090 (24 GB)
5a	Speech-to-Video	S2V-14B	480P / 720P	Image + audio file → talking video
5b	Pose-Driven Speech-to-Video	S2V-14B	480P / 720P	Image + audio + pose MP4
5c	TTS Speech-to-Video	S2V-14B + CosyVoice	480P / 720P	Synthesize voice, then animate
6a	Character Animation	Animate-14B	720P	Character mimics motion from video
6b	Character Replacement	Animate-14B	720P	Swap character into existing video

---

Quick Start

1. Open in Google Colab

Click the badge above or upload wan2_2.ipynb directly to colab.research.google.com.

2. Select a GPU Runtime

Go to Runtime → Change runtime type → select A100 (recommended for A14B models) or T4/L4 (for TI2V-5B).

3. Run Section 0 — Setup

Run the setup cells once per session. They:

• Check your GPU
• Clone the Wan2.2 repo (skips if already present)
• Install all Python dependencies
• Optionally install flash-attn and CosyVoice (S2V TTS) extras

4. Run Section 1 — Download Your Model(s)

Each model has its own download cell. Only download what you need:

Cell	Model	Approx Size
1.1	Wan2.2-T2V-A14B	~28 GB
1.2	Wan2.2-I2V-A14B	~28 GB
1.3	Wan2.2-TI2V-5B	~10 GB
1.4	Wan2.2-S2V-14B	~28 GB
1.5	Wan2.2-Animate-14B	~28 GB

5. Jump to Any Generation Section

Each section is self-contained. Configure settings via the @param widgets, upload your inputs, and run the generation cell.

---

Notebook Structure

wan2_2.ipynb
├── Section 0 — Setup & Installation
│   ├── 0.1  GPU check
│   ├── 0.2  Clone Wan2.2 repo
│   ├── 0.3  Install core dependencies
│   ├── 0.4  Install flash-attn (optional)
│   ├── 0.5  Install S2V / CosyVoice deps (optional)
│   └── 0.6  Install huggingface-hub CLI
│
├── Section 1 — Model Download
│   ├── 1.1  T2V-A14B
│   ├── 1.2  I2V-A14B
│   ├── 1.3  TI2V-5B
│   ├── 1.4  S2V-14B
│   └── 1.5  Animate-14B
│
├── Section 2 — Text-to-Video (T2V-A14B)
│   ├── 2.1  Configuration (resolution, steps, seed, prompt extension)
│   └── 2.2  Run generation
│
├── Section 3 — Image-to-Video (I2V-A14B)
│   ├── 3.1  Upload image
│   ├── 3.2  Configuration
│   └── 3.3  Run generation
│
├── Section 4 — Text+Image-to-Video (TI2V-5B)
│   ├── 4.1  Upload image (optional — omit for pure T2V)
│   ├── 4.2  Configuration
│   └── 4.3  Run generation
│
├── Section 5 — Speech-to-Video (S2V-14B)
│   ├── 5.1  Upload image & audio
│   ├── 5.2a Basic S2V config + run
│   ├── 5.3b Pose-driven S2V (upload pose video) + run
│   └── 5.4c TTS S2V — CosyVoice voice cloning + run
│
├── Section 6 — Character Animation & Replacement (Animate-14B)
│   ├── 6.1  Upload character image & motion video
│   ├── 6.2  Choose mode (animate / replace) & resolution
│   ├── 6.3  Preprocess input video (extracts pose/face signals)
│   ├── 6.4a Run — Animation mode
│   └── 6.4b Run — Replacement mode
│
└── Section 7 — Display & Download
    ├── 7.1  List all generated videos
    ├── 7.2  Preview video inline
    └── 7.3  Download to local machine

---

GPU Requirements

Model	Minimum VRAM	Recommended
TI2V-5B	24 GB	RTX 4090 / L4
T2V-A14B	24 GB (with offload flags)	A100 80 GB
I2V-A14B	24 GB (with offload flags)	A100 80 GB
S2V-14B	80 GB	A100 80 GB
Animate-14B	80 GB	A100 80 GB

All A14B sections include --offload_model, --convert_model_dtype, and --t5_cpu toggle flags to reduce VRAM usage on smaller GPUs.

---

Prompt Extension

For richer, more detailed outputs, the T2V and I2V sections support prompt extension via:

• Local Qwen — runs a Qwen2.5 LLM (T2V) or Qwen2.5-VL (I2V) locally to expand your prompt. No API key needed.
• Dashscope API — uses Alibaba Cloud's hosted qwen-plus / qwen-vl-max models. Requires a free Dashscope API key.

---

Troubleshooting

Out of Memory (OOM)

• Enable all three memory flags: offload_model, convert_model_dtype, t5_cpu
• Drop resolution to 832×480
• Use TI2V-5B (Section 4) instead of the A14B models — it runs on 24 GB

Model Download Fails

• Re-run the download cell — huggingface-cli resumes partial downloads
• Log in to Hugging Face to avoid rate limits: uncomment the login() line in cell 0.6

flash-attn Build Fails

• Skip cell 0.4 — the model falls back to standard attention automatically
• Or try: pip install flash-attn --no-build-isolation after installing all other deps first

S2V / CosyVoice Import Errors

• Make sure you ran cell 0.5 before Section 5c
• CosyVoice requires the requirements_s2v.txt extras

Animate Preprocessing Fails

• Ensure Wan2.2-Animate-14B/process_checkpoint exists (downloaded in cell 1.5)
• Input video should be a standard MP4 with a clearly visible human subject

---

Repository Structure

wan2.2-google-colab/
├── wan2_2.ipynb    # Main notebook (all generation modes)
├── README.md       # This file
└── SECURITY.md

---

Credits & Links

• Wan2.2 Model: Wan-Video/Wan2.2
• Hugging Face: Wan-AI
• Paper: arXiv:2503.20314
• CosyVoice (TTS): FunAudioLLM/CosyVoice

License

This template is released under the same Apache 2.0 License as the upstream Wan2.2 project.

---

This is an unofficial Colab template. For questions about the Wan2.2 model itself, refer to the official repository or join their Discord.