VC-Attention is a training-free low-bit attention framework for video diffusion transformers that tackles both accuracy loss from value outliers and the softmax bottleneck in quantized attention kernels. It combines V-Smooth, which clusters and re-centers value tokens before quantizing only the residual, with ExpCast-FP8, which maps scores directly to FP8 probability codes to remove the FP32 exponential step. Implemented across five GPU generations and tested on four video models, it improves fidelity over existing low-bit baselines and speeds up attention by 1.46-3.6x over BF16 FlashAttention-4, translating into 1.13-1.7x faster end-to-end generation.
