AI video generation has come a long way, and Wan2GP is one of the most accessible options for running on consumer hardware. Unlike cloud-dependent services, Wan2GP runs locally on your own GPUs.
What You Need
Wan2GP requires a GPU with at least 8GB VRAM for basic video generation. For anything beyond short clips (4-8 frames), 12GB+ is recommended.
Minimum specs:
- GPU: 8GB VRAM (RTX 3070 / 4060)
- RAM: 16GB system
- Storage: 10GB free for models
- Python 3.10+
Recommended (dual GPU):
- 2x RTX 3080 Ti (12GB each)
- 32GB system RAM
- CUDA 12.4+
Installation
# Clone the repo
git clone https://github.com/Wan2GP/Wan2GP
cd Wan2GP
# Create environment
python -m venv venv
source venv/bin/activate
# Install with CUDA 12.4 support
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu124
pip install -r requirements.txtOn a dual GPU setup, the model automatically splits across both cards. You'll see each GPU handling around 8-10GB during generation.
Generating Your First Video
Wan2GP uses a diffusers-based pipeline, similar to Stable Diffusion but for video:
from wan2gp import Wan2GPPipeline
import torch
pipeline = Wan2GPPipeline.from_pretrained(
"wan2gp/wan2gp-base",
torch_dtype=torch.float16
).to("cuda")
prompt = "A mountain landscape at sunset, cinematic quality"
video = pipeline(
prompt=prompt,
num_frames=16,
num_inference_steps=50,
guidance_scale=7.5
).frames[0]
video.save("output.mp4", fps=8)Performance Tuning
Frame count vs quality. 8-16 frames is the sweet spot on dual 3080 Ti. Going to 32+ frames requires either reduced resolution or CPU offloading. Each frame adds roughly 1-2 seconds of generation time at 50 steps.
Resolution. 512x512 is fast and reliable. 768x768 works but doubles VRAM usage. For widescreen, use 768x448 to keep the pixel count manageable.
Step count. 30 steps gives decent results. 50 steps for quality. Past 50, you get diminishing returns.
ComfyUI Integration
If you're already using ComfyUI, Wan2GP can be integrated as a custom node. Drop the wan2gp directory into ComfyUI/custom_nodes/ and restart. You'll get a "Wan2GP Video" node that accepts prompts and outputs latent video frames.
Common Issues
Out of memory. Reduce num_frames to 8, lower resolution to 384x384, or enable enable_attention_slicing() on the pipeline.
Slow generation. Enable torch.compile() on the UNet for a 15-20% speedup. First run will be slow (compilation), subsequent runs are faster.
Flickering. Increase num_frames and reduce guidance_scale to 5.0-6.0. Lower CFG reduces temporal artifacts.
Local AI video is still early, but Wan2GP makes it practical on consumer hardware. Start with short clips and experiment your way up.