AIAny
Icon for item

AlayaWorld: Interactive Long-Horizon World Modeling -- Full Technical Report

Generates interactive long-horizon 24-fps video worlds (540p/720p) from text, image, or video inputs. Uses a 15B video diffusion transformer with a bounded visual context (sink frame, compressed temporal history, geometry-aligned spatial memory, recent-frame conditioning) and a discrete autoregressive distillation that cuts inference to ~4 sampling steps per chunk.

Introduction

The core insight: scaling video world models requires not only larger generative backbones but also memory structures and training procedures that prevent long-horizon drift while keeping inference interactive. AlayaWorld targets that gap by pairing a large video diffusion transformer with a compact, persistent visual context and a distillation pipeline that drastically reduces sampling cost.

Key Findings
  • Long-horizon interactive generation: produces 24-fps video at 540p and 720p across extended roll-outs by generating short latent chunks autoregressively under camera trajectories and switchable text prompts, supporting interactive control.
  • Bounded visual context design: combines a persistent sink frame, compressed temporal history, geometry-aligned spatial memory, and recent-frame conditioning to balance context fidelity and compute; this reduces spatial/temporal inconsistency over long sequences.
  • Training for stability: trains on corrupted histories and prediction residuals sampled from the model's own roll-outs to reduce drift and exposure bias in multi-chunk generation.
  • Fast inference via distillation: introduces a discrete autoregressive distillation that blends distribution-matching, self-forcing++, and consistency distillation to cut sampling from roughly 30 steps down to about 4 steps per chunk.
  • Empirical result: achieves top performance on iWorld-Bench for long-horizon video generation benchmarks.
How it works (concise)

AlayaWorld is built on a 15B-parameter video diffusion transformer that outputs short latent chunks. A bounded-context module supplies a persistent reference frame, compressed temporal summaries for long-range state, geometry-aligned spatial memory for consistent object placement, and recent-frame conditioning for local detail. The model is exposed during training to its own roll-outs (with injected corruption) so it learns to correct accumulated errors. For latency-sensitive interactive use, the team distills the sampling process into a discrete autoregressive form that enables high-quality generation in ~4 sampling steps per chunk.

Who it's for and trade-offs

Great fit if you need an open research foundation for interactive, explorable video worlds or want a starting point for building long-horizon control/interaction demos that accept text/image/video prompts. Look elsewhere if you require photorealistic 4K output, extremely low compute-cost clients (the system is heavy at training and relies on a large 15B backbone), or turnkey game-engine integrations—AlayaWorld is presented as a research/full-stack foundation rather than a packaged commercial runtime.

Information

  • Websitearxiv.org
  • AuthorsAlayaWorld Team, Kaipeng Zhang, Chuanhao Li, Yifan Zhan, Yongtao Ge, Yuanyang Yin, Jiaming Tan, Kang He, Liaoyuan Fan, Mingliang Zhai
  • Published date2026/07/20

Categories

More Items

Personalizes subject-driven videos to preserve human identity and accurate human–object interactions by integrating multimodal references and MLLM-derived semantics. Introduces global multimodal guidance in self-attention and modality-reference embeddings to align MLLM features with VAE tokens, supporting both inter- and intra-subject inputs (e.g., OCR, multi-view).

Evaluates whether video models reason according to physical laws by treating generated videos as visible reasoning traces and using a three-stage Perception–Formulation–Deduction protocol. Includes Orchard (400 mechanics videos), chain-of-frames prompting on annotated first frames, and a hybrid MLLM-plus-objective scoring suite for stage-resolved diagnostics.

Predicts variable-cardinality sets of evidence intervals in videos to temporally ground queries using multimodal large language models. Combines caption-derived multi-span supervision, a temporal Wasserstein matching-free reward, and temporal IoU, yielding strong mIoU gains across multiple benchmarks.