SpheRoPE: Zero-Shot Optimization-Free 360 Panorama Generation with Spherical RoPE

2026-06-30

AI Summary

Existing methods for 360-degree panoramic image generation typically rely on dataset-specific fine-tuning or time-consuming optimization, struggling to balance generalization and inference efficiency. This work proposes a zero-shot generation framework that injects spherical priors into pretrained diffusion Transformers, enabling high-quality panoramic image and video synthesis without any fine-tuning or test-time optimization. The core innovations include a spherical RoPE positional encoding scheme, in which low-frequency channels are parameterized as 3D Cartesian coordinates to model the spherical manifold, while high-frequency channels employ harmonic quantization to enforce periodicity. Additionally, semantic distortion-aware classifier-free guidance (CFG) is introduced to enhance geometric consistency. The approach achieves performance on par with or superior to existing baselines across multiple backbone models, including Flux.1, Flux.2, and LTX-Video.

Abstract

We present a zero-shot, training-free and optimization-free framework for generating 360 panoramic images and videos by directly injecting spherical priors into pre-trained diffusion transformers. Existing methods either rely on costly fine-tuning on scarce panoramic data that limits generalization, or leverage multi-step optimization that incurs prohibitive inference latency. We observe that contemporary generative models natively exhibit some panoramic priors from large-scale training. However, these emergent capabilities are insufficient, as the models fundamentally fail to satisfy the rigorous topological constraints imposed by equirectangular projection (ERP). We introduce a zero-shot and optimization-free approach that resolves these constraints at inference time. Spherical RoPE replaces standard rotary position embeddings: low-frequency channels are re-parameterized as 3D Cartesian coordinates to natively encode the spherical manifold, while high-frequency channels are harmonically quantized to enforce exact periodicity. Coupled with complementary Semantic Distortion classifier-free guidance (CFG) that explicitly steers geometry, we avoid retraining and inherit the full creative breadth of state-of-the-art models. Our approach generalizes across diverse backbones and 360 generation modalities. We demonstrate this across text-to-panorama using Flux.1, Flux.2, and LTX-Video backbones, achieving competitive performance against baselines, all while remaining training-free. Project page: https://orhir.github.io/SpheRoPE

Problem

• 360 panorama generation
• equirectangular projection
• topological constraints
• zero-shot generation
• diffusion transformers

Innovation

• Spherical RoPE
• zero-shot generation
• 360 panorama
• optimization-free
• equirectangular projection