Hy-Nomas: Towards Free-form Interactive World Generation
Qiaoling Chen, Jiaxiang Cheng, Xueqing Deng, Tiankai Hang, Ji Hou, Jiaqi Li, Yifan Liu, Tao Lu and 30 others
Tencent Hy
Published October 27, 2026 · on the feed October 28, 2026 · ▲ 123 · arXiv 2610.36729
The abstract
Interactive world generation requires models to follow successive commands while preserving relevant history and action consequences. We present Hy-Nomas, a video world model with an 8B diffusion backbone that supports free-form text instructions and continuous camera control across multiple interaction turns. It generates variable-length segments conditioned on current inputs and cached history. Our central design aligns structured video captions with action-aware causal attention: descriptive context remains accessible across segments, while action instructions are restricted to their intended temporal windows. A vision-language Prompt Enhancer (PE) translates user requests and multimodal interaction histories into structured conditions for the next segment. A multi-stage curriculum progressively introduces causal generation, text interaction, camera control, and training on model-generated histories, followed by diffusion distillation. In human evaluation, Hy-Nomas shows superior performance over existing world models. We release model weights, code and benchmarks, together with a lightweight adaptation workflow, to support reproducible research on interactive world generation.
The Brief
1:47A short podcast summary of the paper. Written and voiced by AI.
Transcript
Hy-Nomas. Towards Free-form Interactive World Generation. By Qiaoling Chen and thirty-seven others, from Tencent HY. Published on the twenty-seventh of October.
Interactive world generation asks a model to follow one command after another, while keeping the relevant history and the consequences of earlier actions intact. Hy-Nomas is a video world model built for that. It sits on an eight billion parameter diffusion backbone, accepts free-form text instructions and continuous camera control across multiple interaction turns, and generates segments of variable length, conditioned on current inputs and on cached history.
The central design pairs structured video captions with action-aware causal attention. The two kinds of conditioning are treated differently. Descriptive context stays accessible across segments, while action instructions are confined to the temporal windows they are meant for. A description of the world persists. A command applies when it applies.
Feeding that structure is a vision-language Prompt Enhancer, or PE. It translates user requests and multimodal interaction histories into structured conditions for the next segment. What a person asks for in free form becomes structured input for the generator. Training follows a multi-stage curriculum that brings in causal generation, then text interaction, then camera control, then training on histories the model itself generated, and ends with diffusion distillation.
Judged by people rather than by an automatic metric, in human evaluation, Hy-Nomas shows superior performance over existing world models. The authors release the model weights and benchmarks, along with a lightweight adaptation workflow, to support reproducible research on interactive world generation.
Keeping memory and commands apart is the organising choice, and the rest of the system is arranged around it. What describes the world is allowed to carry forward, what tells it to change is held to its moment, and the curriculum teaches the model to build on its own past output.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
A world that listens has to remember the right things and forget the commands.
You are building a video model people can play inside: they type commands or move the camera, and the world keeps going, segment after segment, remembering what happened before.
The road ahead
- 1 · The problem ~6 min A playable video world must let a one-off command happen once while the description of the place keeps holding.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Scene text stays visible to later segments, action text is limited to its time window, and training uses the model's own history.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Training runs as a curriculum: causal generation, then text, then camera, then the model's own histories, then distillation.
- 6 · The experiment ~6 min You’ll guess: Judged by people rather than an automatic metric, how did Hy-Nomas compare with existing world models?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.