Selected work

Representative Papers

JoLT: Joint Latent Trajectories for Context-Guided High-Resolution Tiled Generation

Aug 15, 2026

This study addresses the challenge of generating detail-rich, high-resolution images with text-to-image models by proposing a Joint Latent Trajectory method. The approach introduces a dual-stream synchronous denoising mechanism that dynamically integrates low-resolution layouts with high-resolution details through cross-branch information exchange during sampling, combined with a patch-wise generation strategy for coordinated control. Experimental results demonstrate that this method effectively resolves the structural-textural imbalance inherent in high-definition generation, producing visually appealing images with rich details. Significantly outperforming existing baselines, this work establishes a novel paradigm for high-resolution image synthesis.

0 citationsRead paper

Spatially-Grounded Text-to-Video Generation via Inference-Time Gradient-Free Optimization

Aug 13, 2026

Existing text-to-video generation models rely on computationally expensive gradient-based optimization to achieve fine-grained spatial controllability. This work proposes a training-free, gradient-free inference-time optimization method that introduces, for the first time in this task, an analytical trajectory control mechanism. By explicitly solving cross-attention scores within diffusion Transformers and dynamically injecting spatial guidance signals based on the latent space manifold structure, the approach circumvents backpropagation entirely. It achieves substantially improved object localization accuracy over current baselines while incurring only minimal additional computational overhead.

0 citationsRead paper

HyperbolicDiffusion: Sharp & Scalable Tiled Generation on the Hyperbolic Plane

Aug 04, 2026

This work addresses the challenge of image generation in the hyperbolic plane, where the absence of Euclidean-like rectangular canvases and exponential area growth render conventional overlapping-window diffusion methods ineffective. The authors propose the first training-free approach to hyperbolic image synthesis by modeling window placement as a compact dynamic programming problem via Hyperbolic Blooming Cover. A shared implicit canvas is constructed using persistent surface IDs, enabling standard diffusion models to denoise local windows independently before fusing them coherently. To resolve blurriness and inconsistencies at multi-window boundaries, a geometry-aware two-stage re-noising mechanism is introduced. The method produces sharp, view-consistent images that support reprojection, offering a prompt-driven generation framework for artworks in the style of Escher’s *Circle Limit* series.

0 citationsRead paper

Recent publications

Latest Papers

JoLT: Joint Latent Trajectories for Context-Guided High-Resolution Tiled Generation

Aug 15, 2026

This study addresses the challenge of generating detail-rich, high-resolution images with text-to-image models by proposing a Joint Latent Trajectory method. The approach introduces a dual-stream synchronous denoising mechanism that dynamically integrates low-resolution layouts with high-resolution details through cross-branch information exchange during sampling, combined with a patch-wise generation strategy for coordinated control. Experimental results demonstrate that this method effectively resolves the structural-textural imbalance inherent in high-definition generation, producing visually appealing images with rich details. Significantly outperforming existing baselines, this work establishes a novel paradigm for high-resolution image synthesis.

0 citationsRead paper

Spatially-Grounded Text-to-Video Generation via Inference-Time Gradient-Free Optimization

Aug 13, 2026

Existing text-to-video generation models rely on computationally expensive gradient-based optimization to achieve fine-grained spatial controllability. This work proposes a training-free, gradient-free inference-time optimization method that introduces, for the first time in this task, an analytical trajectory control mechanism. By explicitly solving cross-attention scores within diffusion Transformers and dynamically injecting spatial guidance signals based on the latent space manifold structure, the approach circumvents backpropagation entirely. It achieves substantially improved object localization accuracy over current baselines while incurring only minimal additional computational overhead.

0 citationsRead paper

HyperbolicDiffusion: Sharp & Scalable Tiled Generation on the Hyperbolic Plane

Aug 04, 2026

This work addresses the challenge of image generation in the hyperbolic plane, where the absence of Euclidean-like rectangular canvases and exponential area growth render conventional overlapping-window diffusion methods ineffective. The authors propose the first training-free approach to hyperbolic image synthesis by modeling window placement as a compact dynamic programming problem via Hyperbolic Blooming Cover. A shared implicit canvas is constructed using persistent surface IDs, enabling standard diffusion models to denoise local windows independently before fusing them coherently. To resolve blurriness and inconsistencies at multi-window boundaries, a geometry-aware two-stage re-noising mechanism is introduced. The method produces sharp, view-consistent images that support reprojection, offering a prompt-driven generation framework for artworks in the style of Escher’s *Circle Limit* series.

0 citationsRead paper