Meituan's AI livestreamer stack: 82% GTV growth built on four proprietary techniques
美团技术团队 · wechat · 2026-09-03
Meituan details the full tech stack behind its AI digital-human livestreaming product, which grew monthly GTV 82.12% YoY and broadcast sessions 11x. Key components:
- SDIP + SREdit: structure-decoupled identity personalization for 1:1 human replication, plus self-rewarding editing where one model acts as both editor and judge to avoid reward hacking (+2.53.8% on editing benchmarks, ACMMM 2026).
- MoTiGA: causal-LLM-based motion generation; fully causal convolutions cut reconstruction FID from 0.114 to 0.031, with 101,490 human preference pairs for DPO-style optimization (FID 0.041 on HumanML3D).
- StreamingTalk: streaming causal generation of semantically-coordinated gestures with constant latency and memory for unlimited-length real-time performance (submitted to AAAI 2027).
- Glance2Gaze: saccade-inspired visual token compression delivering 75% token reduction and 2.5x inference speedup, enabling tens of thousands of concurrent streams.
A thorough engineering retrospective on taking digital humans from "usable" to mass production.
More from Companies & People
- Reva, the EV that outsold Tesla's peers a decade before Tesla, also died — santoshpanda · 2026-09-03
- Simputer's legal advisor revisits the touchscreen computer that died years before the iPhone — santoshpanda · 2026-09-03
- Musk Reshuffles SpaceX Data Center Team, Moves Rocket Executives to AI Infra — 创业邦 · 2026-09-03
- Sam Altman confirms OpenAI will build a humanoid robot: 'the world is designed for people' — Scobleizer · 2026-09-03
- Meta's billions in token spend fuel prediction that top tech firms will dump Anthropic — tinyfool · 2026-09-03
- GitHub announces first-ever Copilot Day on Sept 10 with demos and product announcements — unixterminal · 2026-09-03