Alibaba TaoLive ships 3-step LoRA bringing real-time streaming audio-video gen to MiniMax H3
linoy_tsaban · x · 2026-10-15
Alibaba's TaoLive AIGC team released TaoMate-H3-FL2VA on Hugging Face, described as the first real-time streaming first/last-frame-to-audio-video (FL2VA) model.
- Built on MiniMax H3 with a 3-step LoRA for blockwise streaming denoising, generating speech, sound, and video on one synchronized timeline
- Supports 480p/768p/1080p resolutions; clean AV-KV caching preserves visual identity, voice, and motion across long-form generations
- Runs single-node on 4/8 GPUs (TP2 + Ulysses), with per-chunk latency well below a full H3 request
- The team migrated H3 to TaoMate (T2AV) in September; full FL2AV and Ref2AV variants are on the way
More from Multimodal
- Open-source ff-tracking hits 407 stars in 6 days as community ports and remixes it — PurzBeats · 2026-10-15
- Purported Gemini 4 builds a shakable New York snow globe entirely from code, zero assets — prasenx · 2026-10-15
- This week in AI film history: Firefly Video launch, Veo 3.1 and more milestones — DavidmComfort · 2026-10-15
- Viral Argon gummy bear video was fake: real Gemini 4 Argon results much better — socialwithaayan · 2026-10-15
- One prompt gets Gemini 4 Argon to build a walkable 3D city in iPhone POV — socialwithaayan · 2026-10-15
- Gemini 4 Argon spawns flood of 3D demos one day after launch: 10 examples — socialwithaayan · 2026-10-15