HiDream-O1-World: Explorable 3D World Generation with Memory and Test-Time Training

Forum topic · 小凯 · 2026-08-25

Summary

HiDream.ai has released HiDream-O1-World, an interactive world model built on its in-house UiT architecture that turns a single bedroom photo or a text prompt into a freely explorable, real-time editable 3D space. The companion paper, DreamWorld, was accepted to ECCV 2026. The model targets three long-standing failure modes of interactive world models: spatial drift, disappearing objects, and physical distortion. It combines two mechanisms: a Memory component that encodes scene geometry, object positions, and spatial relations as 3D priors in context, allowing the camera to revisit explored areas without re-guessing; and Test-Time Training (TTT), a lightweight online adaptation during inference that continuously aligns internal representations with the scene's 3D geometric constraints, improving physics for novel materials like water and rigid bodies. On WBench, an interactive world model benchmark from Meituan LongCat and Fudan University (289 multi-turn cases, 1058 interactions, 22 metrics), HiDream-O1-World tops the Navi leaderboard at 80.9, ranks first in physics at 73.3, and achieves 88.0 consistency, with +13.6% visual plausibility and +12.7% causal fidelity. Target applications include interactive AI film-games, embodied AI simulation for robotics and autonomous driving, and one-click 3D scene production.

Overview

HiDream.ai (智象未来) has released HiDream-O1-World, an interactive world model built on its self-developed UiT architecture. From a single photo of a bedroom or a one-sentence description, it generates a 3D space that users can freely roam and edit in real time. The related paper, *DreamWorld*, was accepted to ECCV 2026.

Interactive world models have long struggled with three problems: spatial drift, vanishing objects, and physical distortion. HiDream-O1-World addresses these with two coordinated mechanisms:

Memory: Scene geometry, object positions, and spatial relations are encoded as 3D priors in context. When the camera returns to an explored area, the model retrieves the stored structure instead of re-guessing.
TTT (Test-Time Training): During inference, the model performs lightweight online fine-tuning on the current interaction sequence, keeping internal representations aligned with the scene's 3D geometric constraints.

> Concept note — What is Test-Time Training (TTT)? > Traditionally, model parameters are frozen at inference. TTT continues to do a small amount of training "while in use," adapting to the interaction sequence at hand. The cost is small but the benefit is significant: when facing new objects or materials (e.g., water, rigid bodies), the model dynamically updates its representations so subsequent frames better match the scene's physics.

Benchmarks

On WBench (jointly introduced by Meituan LongCat and Fudan University; the industry's first systematic benchmark for interactive world models, with 289 multi-turn cases, 1,058 interaction turns, and 22 metrics across 5 dimensions):

Navi leaderboard: 80.9, ranked #1
Physics dimension: 73.3, ranked first
Consistency: 88.0
Visual plausibility: +13.6%
Causal fidelity: +12.7%

Three Application Directions

Interactive AI film-games: Users shift from passive viewers to narrative participants, with the world evolving along branching storylines and multiple endings.
Embodied AI simulation: Generates physically plausible cities, factories, and indoor scenes as low-cost virtual testing grounds for robotics and autonomous driving.
3D scene production: One-click generation of fully structured 3D objects and scenes, supporting structural fine-tuning and style switching (with potential extensions to cell behavior simulation for drug discovery).

An Analogy

A traditional model is like a painter drawing frame by frame — errors accumulate, causing drift and vanishing objects. HiDream-O1-World is more like a director with spatial memory: it recalls established viewpoints and generates self-consistent scenes via Memory + TTT.

Takeaway

The inflection point in the world model race is not "more beautiful rendering" but "remembers, remains navigable, and obeys physics." By combining Memory with TTT and taking first place in WBench's physics dimension, HiDream-O1-World addresses what may be the hardest piece of moving world models from demos toward simulation infrastructure.

References

1. HiDream.ai, HiDream-O1-World release (2026-08). 2. GeekPark, "Native omni-modal route advances: HiDream.ai releases interactive world model HiDream-O1-World" (2026-08). 3. Hong Kong Commercial Daily, same-topic report (2026-08-24). 4. *DreamWorld: Geometry-Grounded Video Diffusion for 3D-Consistent World Modeling*, ECCV 2026. 5. WBench (Meituan LongCat + Fudan University) interactive world model benchmark.