Shengshu's ViduS2 Powers an AI Livestreamer That Earned Over 10K Yuan on Day One
新智元 · wechat · 2026-09-26
A Chinese AI streamer named Zijing, powered by Shengshu Technology's newly released real-time interactive video model ViduS2, held her debut livestream during the Mid-Autumn Festival, drawing 20,000 viewers within 5 minutes. Her challenge: earn 80,000 points (8,000 RMB) in 2 hours or get her memory wiped and restarted. She failed the first round but earned over 10,000 RMB in 5 hours across restarts, improving each round.
The technical highlights are consistency and real-time generation: the SRF training technique gives the model self-correcting memory so the character stays stable through outfit changes, dancing, and viewer interactions; frame-aligned sparse attention, low-bit linear layers, and multi-GPU parallelism cut latency to hundreds of milliseconds. ViduS2 ships as Avatar (real-time voice interaction and motion control) and Editing (real-time restyling of video streams). Shengshu claims top scores on StreamAV-Bench and Sparkle-Bench instruction-following metrics, and points to real-time game NPCs and AI companionship as next applications.
More from Multimodal
- Dev's tested AI music workflow: lyrics, Suno, ear-curation, then Ableton — ctjlewis · 2026-09-27
- Tailored ASR for Japanese speaking assessment cuts mora error rate from 12.3% to 7.1% — tkasasagi · 2026-09-27
- A full AI music video now costs ~$65 and 6M tokens — and it's no longer special — rickasaurus · 2026-09-27
- One Prompt, a 60-Second Singularity Video Essay: Runway CEO Demos Agentic Video Editing — c_valenzuelab · 2026-09-27
- Running Qwen imagegen locally on an 8GB RTX 5060: 'insane results' — tryku2 · 2026-09-27
- Wanted to film a software quality series, made AI songs instead — EricBuess · 2026-09-27