DC-SAE combines semantic and pixel-level encoders to achieve 32x compression with high fidelity and faster diffusion convergence. It achieves 29.79 PSNR and 3.37 gFID on ImageNet 512x512, outperforming prior high-compression tokenizers by large margins.