Selected work

Representative Papers

PE-Field 4D: Video Generation Models as Canvas

Jul 17, 2026

This work addresses the challenge of spatial control in video generation arising from viewpoint changes and camera motion by proposing a geometry-aware diffusion Transformer architecture. By integrating projected positional encoding and a depth-aware disambiguation mechanism, the method effectively fuses 3D depth information with 2D reprojection. It further introduces structured context tokens and geometry-guided cross-attention to enable precise spatial manipulation directly within the native latent space. The proposed approach significantly enhances controllability for viewpoint-dependent editing tasks, supporting camera trajectory redirection, novel view synthesis, and geometry-consistent video editing, all while preserving the strong generative priors of the underlying foundation model.

0 citationsRead paper

Natural Language Camera Movement Understanding

Jul 03, 2026

This work addresses the longstanding challenge that existing vision-language models (VLMs) struggle to accurately interpret camera motion described in natural language, often conflating translation with rotation, left–right directions, and camera versus object motion. To tackle this, the study formally establishes camera motion understanding as a distinct task, introducing a two-tier cinematographic taxonomy and constructing the first atomic-level evaluation benchmark comprising both real-world and synthetic videos. The authors further propose a multi-source data augmentation strategy to train a VLM-8B model. After fine-tuning, the model outperforms Gemini 3.1 Pro by 10% on real videos and 11% on synthetic ones, yet remains substantially below human performance, thereby laying a systematic foundation for future research in this domain.

0 citationsRead paper

Hi-Light: A Path to high-fidelity, high-resolution video relighting with a Novel Evaluation Paradigm

Jan 30, 2026

This work addresses key challenges in video relighting, including the lack of evaluation metrics, temporal flickering, and detail degradation. The authors propose a training-free, high-fidelity, high-resolution video relighting framework that leverages a photometric prior-guided diffusion model for precise illumination control. Temporal flickering is effectively suppressed through a flow-based hybrid motion-adaptive lighting smoothing filter, while high-frequency details are preserved by fusing texture information in the LAB color space. To systematically evaluate lighting consistency, the study introduces the first photometric stability scoring metric for relit videos. Experimental results demonstrate that the proposed method significantly outperforms existing approaches in both qualitative and quantitative assessments, achieving superior lighting stability and rich textural detail.

0 citationsRead paper

UniLayDiff: A Unified Diffusion Transformer for Content-Aware Layout Generation

Dec 09, 2025

Existing approaches struggle to uniformly model heterogeneous constraints—such as element types, dimensions, and spatial relationships—in content-aware layout generation. Method: We propose the first end-to-end, single-model framework capable of unifying unconditional and fully conditional layout generation across the entire spectrum of constraints. Our approach treats layout constraints as independent modalities and introduces a multimodal diffusion Transformer that jointly encodes background images, element semantics, and relational structures. To enable efficient and scalable joint optimization, we incorporate a LoRA-driven relational fine-tuning mechanism. Results: Experiments demonstrate state-of-the-art performance across diverse constraint settings. Notably, our model is the first to comprehensively cover the full task spectrum of content-aware layout generation—spanning unconditional, partially conditional, and fully conditional regimes—thereby establishing a general, flexible, and interpretable foundation for automated graphic design.

0 citationsRead paper

Positional Encoding Field

Oct 23, 2025

Existing Diffusion Transformers (DiTs) rely on 2D positional encodings (PEs), limiting their ability to model 3D geometry and hindering performance in depth-aware tasks such as single-image novel view synthesis and controllable spatial editing. This work identifies the dominant role of PEs in DiT-based visual generation and introduces Positional Encoding Fields (PE-Fields): a structured, 3D PE representation that explicitly encodes depth and hierarchical spatial sub-block relationships, enabling DiTs to directly model volumetric geometry in 3D space. Our method integrates perturbation sensitivity analysis, principled 3D PE design, and a hierarchical encoding mechanism. Experiments demonstrate state-of-the-art performance on single-image novel view synthesis and, for the first time, enable fine-grained, geometrically consistent spatial editing. PE-Fields establish a scalable, geometry-aware encoding paradigm for enhancing the 3D perception capabilities of DiTs.

0 citationsRead paper

Recent publications

Latest Papers

PE-Field 4D: Video Generation Models as Canvas

Jul 17, 2026

This work addresses the challenge of spatial control in video generation arising from viewpoint changes and camera motion by proposing a geometry-aware diffusion Transformer architecture. By integrating projected positional encoding and a depth-aware disambiguation mechanism, the method effectively fuses 3D depth information with 2D reprojection. It further introduces structured context tokens and geometry-guided cross-attention to enable precise spatial manipulation directly within the native latent space. The proposed approach significantly enhances controllability for viewpoint-dependent editing tasks, supporting camera trajectory redirection, novel view synthesis, and geometry-consistent video editing, all while preserving the strong generative priors of the underlying foundation model.

0 citationsRead paper

Natural Language Camera Movement Understanding

Jul 03, 2026

This work addresses the longstanding challenge that existing vision-language models (VLMs) struggle to accurately interpret camera motion described in natural language, often conflating translation with rotation, left–right directions, and camera versus object motion. To tackle this, the study formally establishes camera motion understanding as a distinct task, introducing a two-tier cinematographic taxonomy and constructing the first atomic-level evaluation benchmark comprising both real-world and synthetic videos. The authors further propose a multi-source data augmentation strategy to train a VLM-8B model. After fine-tuning, the model outperforms Gemini 3.1 Pro by 10% on real videos and 11% on synthetic ones, yet remains substantially below human performance, thereby laying a systematic foundation for future research in this domain.

0 citationsRead paper

Hi-Light: A Path to high-fidelity, high-resolution video relighting with a Novel Evaluation Paradigm

Jan 30, 2026

This work addresses key challenges in video relighting, including the lack of evaluation metrics, temporal flickering, and detail degradation. The authors propose a training-free, high-fidelity, high-resolution video relighting framework that leverages a photometric prior-guided diffusion model for precise illumination control. Temporal flickering is effectively suppressed through a flow-based hybrid motion-adaptive lighting smoothing filter, while high-frequency details are preserved by fusing texture information in the LAB color space. To systematically evaluate lighting consistency, the study introduces the first photometric stability scoring metric for relit videos. Experimental results demonstrate that the proposed method significantly outperforms existing approaches in both qualitative and quantitative assessments, achieving superior lighting stability and rich textural detail.

0 citationsRead paper

UniLayDiff: A Unified Diffusion Transformer for Content-Aware Layout Generation

Dec 09, 2025

Existing approaches struggle to uniformly model heterogeneous constraints—such as element types, dimensions, and spatial relationships—in content-aware layout generation. Method: We propose the first end-to-end, single-model framework capable of unifying unconditional and fully conditional layout generation across the entire spectrum of constraints. Our approach treats layout constraints as independent modalities and introduces a multimodal diffusion Transformer that jointly encodes background images, element semantics, and relational structures. To enable efficient and scalable joint optimization, we incorporate a LoRA-driven relational fine-tuning mechanism. Results: Experiments demonstrate state-of-the-art performance across diverse constraint settings. Notably, our model is the first to comprehensively cover the full task spectrum of content-aware layout generation—spanning unconditional, partially conditional, and fully conditional regimes—thereby establishing a general, flexible, and interpretable foundation for automated graphic design.

0 citationsRead paper

Positional Encoding Field

Oct 23, 2025

Existing Diffusion Transformers (DiTs) rely on 2D positional encodings (PEs), limiting their ability to model 3D geometry and hindering performance in depth-aware tasks such as single-image novel view synthesis and controllable spatial editing. This work identifies the dominant role of PEs in DiT-based visual generation and introduces Positional Encoding Fields (PE-Fields): a structured, 3D PE representation that explicitly encodes depth and hierarchical spatial sub-block relationships, enabling DiTs to directly model volumetric geometry in 3D space. Our method integrates perturbation sensitivity analysis, principled 3D PE design, and a hierarchical encoding mechanism. Experiments demonstrate state-of-the-art performance on single-image novel view synthesis and, for the first time, enable fine-grained, geometrically consistent spatial editing. PE-Fields establish a scalable, geometry-aware encoding paradigm for enhancing the 3D perception capabilities of DiTs.

0 citationsRead paper