FastWan-QAD-FP8-1.3B
Introduction
FastWan-QAD-FP8-1.3B is the backward-compatible variant of the FastWan-QAD series, designed for RTX 4090 and other Ampere/Ada GPUs. It uses FP8 quantized linear layers paired with SageAttention2++, generating a 5-second 480p video in approximately 3.4 seconds — still well ahead of prior distilled methods.
The model is built on Wan-AI/Wan2.1-T2V-1.3B-Diffusers and trained with quantization-aware distillation (QAD) for 3-step inference. For RTX 5090 users, see FastWan-QAD-1.3B for maximum speed with NVFP4.
Model Overview
- 3-step inference via quantization-aware distillation
- FP8 linear layers compatible with Ampere, Ada, and Hopper GPUs
- SageAttention2++ backend for attention computation
- Trained at 480p (832×480) resolution, 81 frames (5 seconds at 16 fps)
- No classifier-free guidance at inference time
- Fast decoding via TAEHV tiny autoencoder
Performance
| Model | Hardware | Generation Time (5s 480p) |
|---|---|---|
| FastWan-QAD-1.3B | RTX 5090 | ~1.78s |
| FastWan-QAD-1.3B-SA2 | RTX 5090 | ~2.0s |
| FastWan-QAD-FP8-1.3B | RTX 4090 | ~3.4s |
| TurboDiffusion | RTX 5090 | 6.10s |
| LightX2V | RTX 5090 | 6.91s |
Inference
# Install Tiny Autoencoder
git clone https://github.com/madebyollin/taehv.git
uv pip install -e taehv/
git clone https://github.com/hao-ai-lab/FastVideo.git
cd FastVideo
uv pip install -e .
cd examples/inference/optimizations
python fp8_wan2_1_1_3b.py --taehv-checkpoint /path/to/taehv/taew2_1.pth
Training
More details coming soon.
It would be greatly appreciated if you cite our paper:
@article{Zhang2026AttnQAT,
title={Attn-QAT: 4-Bit Attention With Quantization-Aware Training},
author={Zhang, Peiyuan and Noto, Matthew and Tan, Wenxuan and Jiang, Chengquan and Lin, Will and Zhou, Wei and Zhang, Hao},
journal={arXiv preprint arXiv:2603.00040},
year={2026}
}