FastVideo: Fast Video Generation with Distillation LoRAs

FastVideo provides fast video generation with distillation LoRAs for Wan2.1 and Wan2.2 5B. Enables real-time video generation speeds with fewer sampling steps.

F

FastVideo

Wan2.1Wan2.2DistillationFast InferenceLoRA

Fast video generation framework with distillation LoRAs for Wan2.1 (480p, rank 16/64/128) and Wan2.2 5B full attention model. Enables near real-time video generation speeds.

DeveloperFastVideo (hao-ai-lab)
ArchitectureWan2.1/Wan2.2 with distillation
Model TypeDistillation LoRAs and full model
VariantsWan2.1 distillation LoRAs (480p, rank 16/64/128), Wan2.2 5B full attention
CapabilitiesAccelerated T2V and I2V generation

Overview

FastVideo (also known as FastWan) is a unified inference and post-training framework for accelerated video generation. It provides distillation LoRAs for Wan2.1 at 480p resolution with rank sizes 16, 64, and 128, as well as a Wan2.2 5B full attention model. These distillations enable significantly faster video generation, approaching real-time speeds.

The project also includes FastWan-QAD (Quantization-Aware Distillation), which applies distillation with quantization awareness to achieve even greater speedups while maintaining video quality.

ComfyUI Integration

FastVideo LoRAs are compatible with ComfyUI WanVideo native nodes and Kijai's WanVideoWrapper. Apply the FastVideo distillation LoRA after loading the base Wan2.1 or Wan2.2 model for accelerated generation.

Guides and workflows related to this model series.

No articles found.