Tag
The article benchmarks the acceleration of MiniMax-H3 video generation on 8×H200 GPUs using SGLang Diffusion, achieving up to 6.24× speedup with quality measured by SSIM.
SAC-Copula proposes a quality-preserving watermarking method for diffusion language models using smooth correlated Gumbel fields to improve the trade-off between generation quality and detectability.
This paper proposes a practical evaluation method for long-form simultaneous speech-to-speech translation that uses ASR, forced alignment, and sentence embedding alignment to compute latency and quality metrics on continuous speech, overcoming limitations of prior approaches.