TurboDiffusion:
100-200x Faster Video Diffusion
Generate 5-second videos in under 2 seconds on a single RTX 5090. Powered by SageAttention, SLA, and rCM from Tsinghua University.
From Tsinghua University Machine Learning Group • Read the Paper
Available Models
Four pre-trained models on HuggingFace for Text-to-Video and Image-to-Video generation
TurboWan2.1-T2V-1.3B-480P
Fastest model for quick iterations and prototyping
TurboWan2.1-T2V-14B-480P
Higher quality with 14B parameters
TurboWan2.1-T2V-14B-720P
Best quality for 720p text-to-video generation
TurboWan2.2-I2V-A14B-720P
Transform any image into high-quality video
Generated Examples
Real comparisons between original Wan2.1 and TurboDiffusion-accelerated generation
A stylish woman walks down a Tokyo street filled with warm glowing neon and animated city signage
Dynamic ocean waves crashing against rocky cliffs at sunset
Cinematic aerial view of a futuristic city with flying vehicles
A majestic eagle soaring through mountain peaks in golden hour light
Underwater scene with colorful coral reef and tropical fish
Timelapse of blooming flowers in a spring garden
720P High Resolution Examples (14B Model)
Technical Highlights
Combining three key innovations to achieve unprecedented acceleration
SageAttention Integration
Efficient 8-bit attention with plug-and-play inference acceleration for faster processing
Sparse-Linear Attention (SLA)
Fine-tunable sparse attention that goes beyond traditional sparsity in diffusion transformers
rCM Timestep Distillation
Score-regularized continuous-time consistency for high-quality few-step generation
Single GPU Friendly
Run on consumer hardware like RTX 4090/5090 with quantized checkpoints
Easy Integration
Simple pip install with ComfyUI support. Works with existing Wan model ecosystem
Flexible Quality/Speed
Adjustable parameters for 1-4 step sampling with configurable sigma for quality control
Ready to accelerate your video generation?
TurboDiffusion is fully open source under Apache-2.0 license. Start generating videos 100-200x faster today.