infrastructurefactneutralPost-training quantization of diffusion transformers to W4A4 severely degrades output quality because activations contain outliers that 4-bit formats cannot representMachine Learning29 Jul 2026http://arxiv.org/abs/2607.21446v1