World-Shaper: A Unified Framework for 360° Panoramic Editing

TL;DR

The World-Shaper is presented, a unified geometry-aware framework that bridges panoramic generation and editing within a single editing-centric design, and introduces a geometry-aware learning strategy that explicitly enforces position-aware shape supervision and implicitly internalizes panoramic priors through progressive training.

Abstract

Being able to edit panoramic images is crucial for creating realistic 360° visual experiences. However, existing perspective-based image editing methods fail to model the spatial structure of panoramas. Conventional cube-map decompositions attempt to overcome this problem but inevitably break global consistency due to their mismatch with spherical geometry. Motivated by this insight, we reformulate panoramic editing directly in the equirectangular projection (ERP) domain and present World-Shaper, a unified geometry-aware framework that bridges panoramic generation and editing within a single editing-centric design. To overcome the scarcity of paired data, we adopt a generate-then-edit paradigm, where controllable panoramic generation serves as an auxiliary stage to synthesize diverse paired examples for supervised editing learning. To address geometric distortion, we introduce a geometry-aware learning strategy that explicitly enforces position-aware shape supervision and implicitly internalizes panoramic priors through progressive training. Extensive experiments on our new benchmark, PEBench, demonstrate that our method achieves superior geometric consistency, editing fidelity, and text controllability compared to SOTA methods, enabling coherent and flexible 360° visual world creation with unified editing control. Code, model, and data will be released at our project page: https://world-shaper-project.github.io/

Figures (18)

• Figure 1: Our geometry-aware diffusion framework, World-Shaper, unifies panorama generation and editing within a single editing-centric model, supporting text-to-panorama and image-to-panorama generation, and diverse editing operations such as object addition, removal, replacement, relocation, and appearance modification. New objects and edited regions are highlighted in white and red boxes, respectively.
• Figure 2: Two paradigms for editing panoramic images using existing perspective-based models. First row: applying Nano Banana Pro Sharon2025NanoBanana directly to ERP panoramas suffers from severe spatial distortion. Second row: locally editing cube-map faces and reprojecting them back to ERP reduces distortion but breaks global coherence, causing discontinuities across cube boundaries. Refer to Sec. \ref{['sec:more_motivation']} of Supplemental for more discussions.
• Figure 3: Pipeline of World-Shaper. We adopt a generate-then-edit paradigm: we first train a controllable panoramic generator $G$ to synthesize diverse target panoramas $I_{tgt}$ from source panoramas $I_{src}$ under constraints $C_{gen}$. The synthesized pairs $(I_{src}, P_{\text{edit}}, I_{tgt})$ form a large-scale dataset $D$, on which an editing model $E$ is trained to perform instruction-driven panoramic editing. To address latitude-dependent geometric distortion in ERP images, we introduce a geometry-aware learning strategy that combines Progressive Curriculum Training, which gradually shifting from global panorama generation to localized object manipulation, and Position-aware Shape Constraints, which modulate distortion-aware attention and enforce layered shape consistency across object regions.
• Figure 4: Qualitative comparison with four top-performing SOTA methods from Tab. \ref{['tab:sota']} on panorama editing. White boxes indicate the selected regions for visualization. The corresponding edited areas are shown in the perspective view within the red boxes. Refer to Sec. \ref{['sec:more_results']} of Supplemental for more results.
• Figure 5: Qualitative ablation of Position-Aware Shape Constraints. Editing prompt: "Add a carpet to the right side."