GAF-Pano: Zero-Shot Layout-Controlled Panorama Generation via Global Attention Fusion

ICLR 2026 Conference Submission12701 Authors

Keywords: panorama image generation, layout controlled generation

TL;DR: We propose a new image generation method to generate panoramas with layout control.

Abstract:

Achieving both global semantic coherence and precise local layout control in wide-aspect-ratio panorama generation is an unresolved challenge with potential applications. Existing methods that synchronize independent views to generate panoramas often lack semantic coherence and struggle with fine-grained object placement, resulting in contextual artifacts and fragmented objects.
We introduce GAF-Pano, a training-free framework for zero-shot layout-controlled panorama generation.
GAF-Pano integrates a Global Attention Fusion mechanism into a pre-trained layout-to-image model. Through a Global Context Synchronization, Fusion, and Dispatch workflow, it periodically aggregates latent features from all local views to construct a unified global context, performs multi-level attention computation over this context to achieve true fusion, and then dispatches the enriched global features back to each view, enabling coherent rendering of complex, holistic layouts.
Furthermore, we introduce a conditional positional mask to resolve object repetition artifacts that often arise in large specified regions.
On a newly constructed yet challenging benchmark for panoramic layout control, GAF-Pano achieves superior performance in both layout fidelity and semantic coherence, faithfully generating complex panoramic scenes.

Primary Area: generative models

Submission Number: 12701