PulseQuant: Propagation-Guided Subspace Correction for 4-Bit Video Diffusion Transformers

Yutong Wang1Xingtong Ge2Enhuai Liu1Yunke Wang1Tianfan Xue4,3Xinyuan Chen3Chang Xu1

1The University of Sydney2The Hong Kong University of Science and Technology3Shanghai AI Laboratory4The Chinese University of Hong Kong

4-bit weights · 4/6-bit activations · Post-training quantization

Abstract

Quantization errors in video diffusion transformers can be amplified or attenuated by subsequent denoising updates, making local reconstruction error an incomplete predictor of final impact. We introduce PulseQuant, a 4-bit post-training quantization method that combines trajectory sensitivity with activation geometry to guide offline calibration. Isolated block–step interventions estimate propagation risk, which prioritizes sensitive trajectory states during row-radius selection. With these radii fixed, response-subspace correction uses neighboring-code edits to reduce residual components along dominant activation directions. Both stages preserve the original 4-bit weight representation. Controlled interventions show that short-horizon propagated error predicts final latent error more reliably than immediate block-output error, supporting calibration beyond local reconstruction objectives. Evaluations on Wan models, Self Forcing, and MiniMax-H3 demonstrate improvements in key consistency and dense-reference metrics while remaining competitive on other attributes across model scales and generation paradigms.

Follow the denoising dynamics

Account for how later denoising steps amplify or attenuate quantization errors.

Correct the response subspace

Refine 4-bit weights in activation directions that matter to layer outputs.

Transfer across video models

Explore results on Wan 2.1, Wan 2.2, MiniMax-H3, and Self-Forcing.

PulseQuant overview: propagation-guided radius calibration followed by response-subspace correction.
Figure 2. PulseQuant overview. Propagation-guided radius calibration allocates correction effort according to downstream sensitivity. Response-subspace correction then refines quantized weights along activation directions.

Models & downloads

Download a calibration checkpoint for your model and precision, then use it with the original model weights.

Model4-bit activations6-bit activations
Wan 2.1 · 1.3BW4A4 ↗W4A6 ↗
Wan 2.1 · 14BW4A4 ↗W4A6 ↗
Wan 2.2 · A14BW4A4 ↗W4A6 ↗
MiniMax-H3W4A4 ↗W4A6 ↗
Self-ForcingCalibrate and run with the code ↗

Citation

@article{wang2026pulsequant,
  title={PulseQuant: Propagation-Guided Subspace Correction
         for 4-Bit Video Diffusion Transformers},
  author={Wang, Yutong and Ge, Xingtong and Liu, Enhuai and
          Wang, Yunke and Xue, Tianfan and Chen, Xinyuan and Xu, Chang},
  journal={arXiv preprint arXiv:2609.33384},
  year={2026}
}