Recent advances in audio-driven human video synthesis have significantly improved the realism of half-body and full-body generation. However, generating long-duration sequences remains a persistent challenge, as existing methods often suffer from error accumulation and identity drift over time. While reference image injection strategies (e.g., InfiniteTalk) have been proposed to mitigate these issues, they frequently result in a rigid “copy-paste” phenomenon and limited motion diversity due to conditional image leakage. Furthermore, current models exhibit an undesirable over-reliance on speech signals, leading to unnaturally static behaviors during silent segments. To address these limitations, we present LongCat-Video-Avatar, a unified architecture designed for super-realistic, lip-synchronized long video generation with natural dynamics and consistent identity. It supports multiple video generation modes, including text-audio-to-video, text-audio-image-to-video, and audio-conditioned video continuation. We first analyze the coupling between speech and human motion, proposing a Disentangled Unconditional Guidance strategy that separates audio signals from motion dynamics to ensure natural behavior even in the absence of speech. To alleviate the “copy- paste” issue, we introduce a Reference Skip Attention mechanism that strategically incorporates reference cues to preserve identity while preventing excessive leakage, thereby balancing visual fidelity with motion richness. Additionally, to tackle error accumulation caused by redundant VAE decode-encode cycles in autoregressive generation, we propose a Cross-Chunk Latent Stitching strategy. Extensive evaluations demonstrate the effectiveness of our approach in generating super realistic, long-duration human videos.