A video generation AI model called 'FastWan-QAD' has appeared that can generate videos in 1.8 seconds using an RTX 5090, achieving a 95.6x speed increase by quantizing and distilling Wan2.1.
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
https://haoailab.com/blogs/fastwan-qad/
(1/5) 5 seconds of video. 1.8s seconds of generation. One NVIDIA GeForce RTX 5090 on FastVideo. 🤯🚀
— Hao AI Lab (@haoailab) June 23, 2026
- FastWan-QAD, a new family of video generation models
- Trained with FastVideo's Quantization-Aware Distillation (QAD) recipe.
- Powered by FastVideo, we push a single NVIDIA… pic.twitter.com/zsyJiIEf3A
AI models can be sped up and memory usage reduced by applying a process called 'quantization.' However, quantization is a mechanism that 'reduces computational precision,' and the lower the computational precision, the lower the quality of the generated product. FastWan-QAD is a model that applies a distillation method called Quantization-Aware Distillation (QAD) , which takes quantization into account, to Wan2.1-T2V-1.3B , and has succeeded in speeding up while suppressing the degradation of the generated video quality.
FastWan-QAD is available in three versions: 'FastWan-QAD-1.3B,' 'FastWan-QAD-1.3B-SA2,' and 'FastWan-QAD-FP8-1.3B.' Their respective features are as follows.
FastWan-QAD-1.3B: A model quantized to NVFP4 precision. Inference processing is performed using the SageAttention3 FP4 backend. On an RTX 5090, it can generate a 5-second 480p video in 1.8 seconds.
FastWan-QAD-1.3B-SA2: Performs inference processing with SageAttention2++. Higher quality than FastWan-QAD-1.3B. Can generate a 5-second 480p video in 2.01 seconds on an RTX 5090.
FastWan-QAD-FP8-1.3B: A model quantized to FP8 precision. Because it uses FP8 precision, it can run on RTX 40 series cards. On an RTX 4090, it can generate a 5-second 480p video in 3.4 seconds.
The graph below compares the inference speeds of 'Wan2.1-T2V-1.3B (base model)', 'TurboDiffusion TurboWan', 'LightX2V Wan-NVFP4', 'FastWan-QAD-1.3B-SA2', and 'FastWan-QAD-1.3B'. FastWan-QAD-1.3B-SA2 can perform inference 5.6 times faster than the base model, and FastWan-QAD-1.3B can perform inference 6.4 times faster than the base model.
The graph below compares the time required to generate a 5-second 480p video. FastWan-QAD-1.3B achieves an extremely fast generation speed, 95.6 times faster than the base model.
'FastWan-QAD-1.3B', 'FastWan-QAD-1.3B-SA2', and 'FastWan-QAD-FP8-1.3B' are available at the following link. The license is the Apache License 2.0 .
FastVideo/FastWan-QAD-1.3B · Hugging Face
https://huggingface.co/FastVideo/FastWan-QAD-1.3B
FastVideo/FastWan-QAD-1.3B-SA2 · Hugging Face
https://huggingface.co/FastVideo/FastWan-QAD-1.3B-SA2
FastVideo/FastWan-QAD-FP8-1.3B · Hugging Face
https://huggingface.co/FastVideo/FastWan-QAD-FP8-1.3B
Related Posts:
in AI, Posted by log1o_hf