FastWan-QAD: FastVideo generates a 5-Second Video in 1.8 Seconds on a Single NVIDIA GeForce RTX 5090 via Quantization-Aware Distillation | Hao AI Lab @ UCSD

Generating a 5-second 480P video end-to-end takes just 1.8 seconds on a single NVIDIA GeForce RTX 5090!

UCSD's Hao AI Lab has officially open-sourced FastWan-QAD, a family of distilled video generation models that significantly outperform existing solutions like TurboDiffusion and LightX2V. Built on Wan2.1-T2V-1.3B, this project leverages a novel Quantization-Aware Distillation (QAD) technique to push consumer-grade GPU inference speed to its absolute physical limit.

🔹 **Optimization Stack**: Extreme quantization combined with dense attention. Every major linear layer in the DiT network is quantized to hardware-specific low-bit representations (NVFP4 or FP8) with on-the-fly activation quantization. Unlike other acceleration methods, it completely avoids sparse routing, maintaining 100% dense attention via SageAttention3 FP4 or SageAttention2++ backends to guarantee high video quality.

🔹 **Released Checkpoints**: Three distilled model versions have been published. These include the flagship NVFP4 version optimized for RTX 5090 (1.8s generation), a higher-quality variant utilizing SageAttention2++ (2.0s), and an FP8 version designed specifically to support the previous-generation RTX 4090 architecture.

🔹 **License and Code**: All resources, including model weights, training scripts, and inference code, are released under the Apache-2.0 license. This release drastically lowers the technical and cost barriers for developers running high-speed, local video generation pipelines on consumer hardware.

https://haoailab.com/blogs/fastwan-qad/