Murea Labs' InWorld: AI-Generated World That Remembers Characters and Objects Turns Later
jiqizhixin · x · 2026-10-27
Murea Labs introduced InWorld, coupling a world-planning agent with fast audiovisual generation so the generated interactive world remembers characters, locations, objects and story progress over many turns.
The core Murea-Mind component keeps an explicit record of characters, objects, locations and plot progress, converts user actions into generation instructions, checks generated frames and updates selected facts. Reference images anchor recurring identities, separate audio/video histories continue motion and unfinished speech, and branching histories support alternative storylines.
On benchmarks, Murea-Pro scores 81.13 on WBench Full-289, highest among compared systems; Murea-Fast scores 64.98 on OpenS2V-Eval, leading evaluated open-source models. On eight H20 GPUs, cumulative inference optimizations cut generation time for a 5-second clip from 231 seconds (post text truncated).
More from Multimodal
- MatAnyone-E: 7.4× faster 4K video matting with just 7.6% of peak GPU memory — ccloy · 2026-10-27
- Microsoft's MAI-Image claims #2 spot in image editing, beating Google, Meta and xAI — mustafasuleyman · 2026-10-27
- Avatar Arena launches: face-off arena for 8 AI avatar video models — ycombinator · 2026-10-27
- Microsoft's MAI-Image-2.7-Preview takes #2 on image editing arena, closing in on GPT Image 2.5 — ArtificialAnlys · 2026-10-27
- SenseNova 6.8 Code2Video takes the cycling pelican all the way to the seaside — Single-Climate2259 · 2026-10-27
- SAM 3 author praises single-camera 3D rebuild of Curry's iconic shot using SAM models — nikhilaravi · 2026-10-27