Video World Models · 2026-10-09
IntactWorld: Joint World Modeling with Intact Features
Boming Tan, Xiangdong Zhang, Yan Xia, Qi Zhu, Deyi Ji, Xue Yang, Shaofeng Zhang
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
IntactWorld is a joint video world-modeling architecture that uses “intact” (uncompressed) semantic/spatial/temporal features. To avoid a manifold-gap bottleneck, it predicts clean world features x0 at intermediate layers while keeping velocity prediction for VAE latents, computing the training loss in velocity space. It also introduces a Full-to-Compact training and compact CLS inner-guidance to reduce inference overhead.
Why it matters
The authors argue that using compressed features loses structural information, while training directly on uncompressed full features causes an optimization “manifold gap” when predicting off-manifold flow velocity in high-dimensional dense feature spaces. IntactWorld targets this with an on-manifold prediction target (clean features) plus velocity-space loss, and then reduces inference cost by abstracting world knowledge into compact CLS tokens for guidance.
Method
- Stage I: joint training with uncompressed dense world features (from DINOv2 and VGGT), predicting clean world features x0 and computing the regression loss in velocity space; velocity prediction remains for VAE latents.
- Stage II: replace dense spatial/semantic/world features with compact CLS tokens and concatenate tokens with VAE latents; use position-isolated RoPE to avoid positional interference.
- Compact Inner-Guidance: extend classifier-free guidance with a single additional CLS branch, reducing forward-pass count from five to three per denoising step versus DreamWorld’s multi-source inner-guidance.
Abstract (from arXiv)
While recent video generation models synthesize highly realistic visuals, they lack a genuine understanding of intrinsic real-world logic. Existing methods attempt to understand the world by internalizing diverse world knowledge, yet constrained by computational overhead or dimensionality alignment, their learning processes inevitably compress features, causing a severe loss of structural information. To address this, we propose \textbf{IntactWorld}, a \textbf{Joint World Modeling Architecture} utilizing uncompressed \textbf{Intact Features}. Since data naturally reside on a low-dimensional manifold within a high-dimensional space, predicting the flow velocity $v$ within this uncompressed high-dimensional space induces a severe manifold gap. To successfully eliminate this optimization bottleneck, our framework instead predicts the clean feature $x_0$ at intermediate layers. Furthermore, to mitigate the computational overhead of incorporating complete world knowledge, we introduce a \textit{Full-to-Compact Training Paradigm}. By replacing raw full features with highly refined CLS tokens, this paradigm enables efficient single-branch guidance, reducing spatial memory consumption by 11.4\% and cutting inference latency by 43.8\%. Extensive evaluations demonstrate the effectiveness of IntactWorld, outperforming established baselines by 2.46 points on the VBench 2.0 benchmark.
Related papers
- World-T3: Test-Time Training as Persistent Memory for Real-Time Video World Models
- WorldFlow: Joint Video and 3D Motion Prediction via Large-Scale Pretraining
- WorldProbe: When Video World Models Fall Apart under Long-Horizon Spatial Exploration
- WorldFoundry: Towards Unified Infrastructure for World Intelligence
- World Embedding Benchmark
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-08 · Indexed by Pub-AI: 2026-10-09 UTC
IntactWorld: Joint World Modeling with Intact Features
Preprint posted: 2026-10-08 · Original record ↗
- arxiv
- 2610.11174
Explicit source versions (1)
- v1 · 2026-10-08
- Preprint available
Source date: 2026-10-08 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
While recent video generation models synthesize highly realistic visuals, they lack a genuine understanding of intrinsic real-world logic. Existing methods attempt to understand the world by internalizing diverse world knowledge, yet constrained by computational overhead or dimensionality alignment, their learning processes inevitably compress features, causing a severe loss of structural information. To address this, we propose \textbf{IntactWorld}, a \textbf{Joint World Modeling Architecture} utilizing uncompressed \textbf{Intact Features}. Since data naturally reside on a low-dimensional manifold within a high-dimensional space, predicting the flow velocity $v$ within this uncompressed high-dimensional space induces a severe manifold gap. To successfully eliminate this optimization bottleneck, our framework instead predicts the clean feature $x_0$ at intermediate layers. Furthermore, to mitigate the computational overhead of incorporating complete world knowledge, we introduce a \textit{Full-to-Compact Training Paradigm}. By replacing raw full features with highly refined CLS tokens, this paradigm enables efficient single-branch guidance, reducing spatial memory consumption by 11.4\% and cutting inference latency by 43.8\%. Extensive evaluations demonstrate the effectiveness of IntactWorld, outperforming established baselines by 2.46 points on the VBench 2.0 benchmark.
