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

TL;DR

A zero-shot, training-free framework that generates 360 panoramic images and videos by injecting spherical position encodings into pre-trained diffusion transformers.

Problem

Existing 360 panorama generation methods rely on costly fine-tuning on scarce data or multi-step optimization that causes high inference latency. Current models also fail to satisfy the strict topological constraints imposed by equirectangular projection (ERP).

Approach

SpheRoPE replaces standard rotary position embeddings (RoPE) with Spherical RoPE. Low-frequency channels are re-parameterized as 3D Cartesian coordinates to encode the spherical manifold, while high-frequency channels are harmonically quantized to enforce exact periodicity. This is coupled with a complementary Semantic Distortion classifier-free guidance (CFG) to explicitly steer geometry, avoiding retraining.

Results & Contribution

The approach achieves competitive performance against baselines in text-to-panorama generation using Flux.1, Flux.2, and LTX-Video backbones, all while remaining training-free. It demonstrates generalization across diverse backbones and 360 generation modalities.