Jingtao Xu, Zizhuo Lin, Jianwen Sun, Yi Yang, Yawei Luo
原文摘要
Localizing targets in $360^\circ$ environments requires combining global context with fine-grained evidence despite projection distortion and horizontal wrap-around. Uneven directional coverage compounds these challenges, particularly for seam-adjacent and polar targets. We present SPIRAL-360, a geometry-aware framework with separately trained single-turn and multi-turn models for panoramic localization. Spherically Stratified Replay (SSR) broadens directional coverage through stratified sampling and consistent transformations of panoramas, target annotations, and tool trajectories. TapSphere incorporates spherical distance priors into panoramic visual attention, while Spherical Direction Regularization (SDR) complements coordinate-token prediction with auxiliary direction supervision during training. The single-turn model predicts directly from the panorama, while the multi-turn model inspects tool-generated perspective views. We construct SPIRAL-360 Search, a curated dataset of trajectories interleaving reasoning, projection actions, and visual observations. Both models undergo full-parameter supervised fine-tuning followed by policy optimization. Experiments demonstrate state-of-the-art localization on S360 and competitive cross-dataset transfer, including H$^*$Bench. Systematic ablations support the effectiveness of this geometric learning design.
Linwei Dong, Ruoyu Guo, Ge Bai, Quan Zheng, Yawei Luo, Changqing Zou
原文摘要
Diffusion distillation, exemplified by Distribution Matching Distillation (DMD), has shown great promise in few-step generation but often sacrifices quality for sampling speed. While integrating Reinforcement Learning (RL) into distillation offers potential, a naive fusion of these two objectives relies on suboptimal raw sample evaluation. This sample-based scoring creates inherent conflicts with the distillation trajectory and produces unreliable rewards due to the noisy nature of early-stage generation. To overcome these limitations, we propose GDMD, a novel framework that redefines the reward mechanism by prioritizing distillation gradients over raw pixel outputs as the primary signal for optimization. By reinterpreting the DMD gradients as implicit target tensors, our framework enables existing reward models to directly evaluate the quality of distillation updates. This gradient-level guidance functions as an adaptive weighting that synchronizes the RL policy with the distillation objective, effectively neutralizing optimization divergence. Empirical results show that GDMD sets a new SOTA for few-step generation. Specifically, our 4-step models outperform the quality of their multi-step teacher and substantially exceed previous DMDR results in GenEval and human-preference metrics, exhibiting strong scalability potential.