Hy-Nomas: Toward Free-Form Interactive World Generation

Problem

An interactive world model has to follow user commands that arrive during generation and maintain the consequences of earlier actions and the scene history. Simply putting a causal mask on a bidirectional DiT is not enough. Errors build up as self-generated outputs accumulate, and it is also hard to separate scene information that should persist from action instructions that apply only at specific times.

Method

Built on HunyuanVideo 1.5, the authors created an 8B MMDiT with segment-level causal attention and camera cross-attention. Training videos were annotated with structured JSON captions containing subject IDs, scene changes and timestamped actions, along with camera Trajectories estimated with ViPE. Action-aware causal attention keeps scene description text visible to later segments, while action text applies only within its time interval. A VLM Prompt Enhancer converts the user's request and history into a structured prompt and length for the next segment. After a multi-stage curriculum of causal generation → text interaction → camera control → training on model-generated history (segment forcing), diffusion Distillation was applied.

Results

• The authors report that Hy-Nomas outperformed existing world models in human evaluation. The portion of the paper provided does not include specific scores.
• The authors state that it supports minute-long Rollouts that mix free-text instructions and camera control within a single session.
• By combining variable-length video packing, an attention kernel that handles heterogeneous masks and distributed sharding, it was trained on packed sequences of up to 360K tokens.
• The authors released the model weights, code, a benchmark and a lightweight adaptation workflow that transfers LoRAs for the base model to the Distilled model.

Limitations

In the provided text, quantitative evaluation figures and details of the comparison baselines are cut off, so the size of the advantage is hard to verify. The authors state that temporal causal structure alone does not guarantee physical causality, and action-consequence learning depends on data and evaluation. The authors also mention remaining challenges, including accumulated error from self-generated history and spatial memory.

Why it matters

As an open-weight world model that lets users keep directing not only camera movement but also 'what happens next' in free text, it provides a foundation that can be used directly in research on games, simulation and interactive content. Its approach of designing caption structure and attention masks together is also a useful reference for other long-video generation research.

Key terms

World model

An interactive virtual environment model that generates the visual observation at the next time step in response to user actions or commands.

Action-aware causal attention

An attention mask design that lets later segments keep referring to scene descriptions, while restricting action instructions to be visible only within a set time interval.

KV cache-based autoregressive generation

An approach that stores the keys/values of previously generated segments and reuses them as history when generating the next segment.

Segment forcing

A training technique that conditions on previous segments generated by the model itself instead of ground-truth video, preparing the model for errors that accumulate at inference time.

Original

• arXiv abstract
• PDF
• HTML full text
• Code
• Project
• Hugging Face

Authors: Qiaoling Chen, Jiaxiang Cheng, Xueqing Deng, Tiankai Hang, Ji Hou, Jiaqi Li, Yifan Liu, Tao Lu, et al. · Posted 2026-10-27 · License arXiv 비독점 배포 라이선스 · Based on the full text

This review is an AI-written summary of the paper. Please treat the original paper as the reference.