Alaya-EVOKE: a video world model that generates 120 minutes straight on a single GPU
量子位 · wechat · 2026-08-17
AlayaLab, USTC and Shanghai AI Lab released Alaya-EVOKE, a video world model achieving hour-long continuous generation: a demo runs 120 minutes, 4,800 clips, 172,800 frames in one take, with stable brightness/saturation/contrast, using only 3 denoising steps and no CFG—2.11s of diffusion for 1.5s of 384×640 video on one H200.
Core idea: don't force memory and real-time into the same denoiser. The student keeps only 3.2s of context; older content is written to an external "world state bank" indexed by camera pose (a persistent point cloud built from monocular depth, read back via top-8 z-buffer warps, evicted on CPU). The teacher is rebuilt on Wan2.2A14B with chunked linear attention, supporting prompt switching mid-generation and long-horizon supervision; distribution matching (DMD) over 30s rollouts transfers this to the 3-step student.
Results: tops WBench navigation subset (80.8 overall as of Aug 16), VBench-2.0 66.77 (above Veo3's 66.72) among public models, though not step-matched. Scene identity similarity (0.523) matches real footage; revisit quality gains 2.3–3.2dB but plateaus at 15.4–17.8dB—"recognizable", not pixel-exact. Mid-generation commands succeed 67% for unanchored regions but only 4% when overriding rendered geometry. Limitations: only static geometry; object identity and fine detail consistency remain limited. Code, weights and project page are public.
Related event: Alaya-EVOKE: 14B World Model Generates 120-Minute Videos on One GPU(2 posts)→
More from Multimodal
- FrankenSim builds a physics-based virtual orchestra, piano and drum demos live — doodlestein · 2026-10-16
- vLLM officially supports Muse Spark Open: multimodal, 1M context, agentic — vllm_project · 2026-10-16
- InfoVis: 9B unified model tops open-source slide generation and editing, trained on 13M records — Jiahao Wang · 2026-10-16
- RepForcing: feature-space supervision improves few-step autoregressive video distillation — zju · 2026-10-16
- SCOPE: semantic span-aware GRPO objective improves T2I prompt rewriting — university-of-rochester · 2026-10-16
- Orbia benchmark asks: do generated videos form a consistent 3D world? — tencent · 2026-10-16