<ul><li>Researchers propose a method for fully quantized Winograd convolution to reduce computational and storage costs in large-scale text-to-image diffusion models.</li><li>Quantization of diffusion models has been explored in previous works to reduce compute costs and memory bandwidth usage.</li><li>The proposed method focuses on finer-grained group-wise quantization, combined with finetuning the scale parameters of the Winograd transform matrices.</li><li>The method achieves near-lossless quality in text-to-image generation and outperforms state-of-the-art methods in image classification tasks.</li></ul>

Data-Free Group-Wise Fully Quantized Winograd Convolution via Learnable Scales

Discover more