Stable Video Diffusion (SVD) 1.1 is Stability AI's open-source video generation model, enabling high-quality video synthesis from image and text inputs.

Model Overview

Key Capabilities

• Image-to-Video: Generate 2-4 second video clips from a single image
• Text-to-Video: Create videos from text descriptions (with XT model)
• Open Source: Full model weights available for local deployment
• Customizable: Fine-tune on custom datasets for specific styles
• Resolution: Up to 1024x576 native, upscalable

Model Variants

• SVD: Base model, 14 frames at 576x1024
• SVD-XT: Extended model, 25 frames at 576x1024
• SVD-XT 1.1: Improved quality, better motion coherence

Local Setup

Requirements

# Hardware requirements
# - NVIDIA GPU with 12GB+ VRAM (RTX 3060 minimum)
# - 32GB RAM recommended
# - 20GB disk space for model weights

# Install dependencies
pip install diffusers transformers accelerate torch
pip install opencv-python pillow

# Download model
from huggingface_hub import snapshot_download
snapshot_download("stabilityai/stable-video-diffusion-img2vid-xt-1-1")

Basic Image-to-Video Generation

import torch
from diffusers import StableVideoDiffusionPipeline
from PIL import Image

pipe = StableVideoDiffusionPipeline.from_pretrained(
    "stabilityai/stable-video-diffusion-img2vid-xt-1-1",
    torch_dtype=torch.float16,
    variant="fp16"
)
pipe.to("cuda")

# Load conditioning image
image = Image.open("input_image.png").resize((1024, 576))

# Generate video
frames = pipe(
    image,
    num_frames=25,
    decode_chunk_size=8,
    motion_bucket_id=127,
    noise_aug_strength=0.02
).frames[0]

# Save as video
from diffusers.utils import export_to_video
export_to_video(frames, "output.mp4", fps=6)

Key Parameters

• motion_bucket_id (1-255): Controls amount of motion (higher = more motion)
• noise_aug_strength (0-1): Noise added to conditioning image
• num_frames: Number of output frames (14 for SVD, 25 for XT)
• decode_chunk_size: Memory optimization parameter