KroQuant: Kronecker-Structured Block Transforms for Efficient Post-Training Quantization of Diffusion Transformers

TL;DR - KroQuant enables higher-quality W4A4 post-training quantization of diffusion transformers using efficient Kronecker-structured activation transforms. It reduces inference overhead while keeping generated outputs closer to full-precision references.

• Applies learned invertible transforms to 32-element activation blocks using small tensor-core GEMMs.
• Stores less than half the parameters of per-channel scaling and runs up to 14% faster than SmoothQuant’s kernel on an MI350 GPU.
• Uses offline LoRaQ calibration to absorb residual per-weight quantization errors.
• Outperforms SVDQuant and LoRaQ in similarity to FP outputs across PixArt-Σ, SANA, and FLUX.1-schnell benchmarks while preserving or improving image quality.