Diffusion generative models have become the standard for producing high-quality, coherent video content, yet their slow inference speeds and high computational demands hinder practical deployment. Although both quantization and sparsity can independently accelerate inference while maintaining generation quality, naively combining these techniques in existing training-free approaches leads to significant performance degradation, as they fail to achieve proper joint optimization. We introduce FPSAttention, a novel training-aware co-design of FP8 quantization and Sparsity for video generation, with a focus on the 3D bi-directional attention mechanism. Our approach features three key innovations: 1) A unified 3D tile-wise granularity that simultaneously supports both quantization and sparsity. 2) A denoising step-aware strategy that adapts to the noise schedule, addressing the strong correlation between quantization/sparsity errors and denoising steps. 3) A native, hardware-friendly kernel that leverages FlashAttention and is implemented with optimized Hopper architecture features, enabling highly efficient execution. Trained on Wan2.1's 1.3B and 14B models and evaluated on the vBench benchmark, FPSAttention achieves a 7.09$\times$ kernel speedup for attention operations and a 4.96$\times$ end-to-end speedup for video generation compared to the BF16 baseline at 720p resolution—without sacrificing generation quality.

Analog transmission schemes for wireless video multicast, especially SoftCast, have gained great attention recently. SoftCast can overcome the “cliff effect” by delivering the linear transformations of video pixels, which is unavoidable for digital schemes. Considering the fact that different parts of image/video have diverse importances to the human perception, in this paper, a new analog video multicast scheme named SCAST is proposed. SCAST is based on SoftCast and the image segmentation technology, designed to improve the video subjective visual quality. Firstly, SCAST employs the Otsu segmentation algorithm to decompose the video source into the foreground part and the background part. Then a power allocation method is presented to provide strong protection for the foreground part, which is the region of interest (ROI) of video for the human visual system (HVS). The two parts are both encoded in the analog way by SoftCast and mapped into a complex signal before transmitted. Results show that the proposed scheme SCAST outperforms SoftCast, H.264/AVC and WSVC in both the metrics, i.e., peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), which reflect the objective and subjective video qualities separately. Especially, SCAST performs excellently under bad channel condition.

Hybrid Digital-Analog (HDA) transmission for wireless video multicast, has gained great attention recently, as it can integrate the advantages of digital and analog coding simultaneously. In this paper, a new HDA video transmission scheme is proposed, named VCAST, which exploits the visual information to improve both subjective visual and objective quality. Since the visual information reflects human perception and mainly exists in middle and high frequency bands, VCAST extracts these bands of video and transmits them in digital way, i.e., H.264/AVC, for high transmission accuracy and good subjective quality. Then the remaining of the video, namely the insensitive information, is processed by 2D-DWT and divided into the approximation and the details parts, i.e., LL band and HL, LH, HH bands. The LL band is also transmitted by H.264/AVC. The HL, LH, HH bands and the residual of LL band are transmitted in analog way by SoftCast, for saving energy and good objective quality. Results show that the proposed VCAST significantly outperforms WSVC, SoftCast and H.264/AVC in both the objective metric (i.e., peak signal-to-noise ratio, PSNR) and the subjective metric (i.e., structural similarity, SSIM).