When zooming between dual cameras on smartphones, prominent discontinuities in both geometric perspective and color rendition inevitably degrade the visual experience of users. To address this, we introduce the dual-camera smooth zoom (DCSZ) task, aiming to synthesize continuous intermediate frames for fluid zooming. However, directly applying existing video frame interpolation (VFI) models to this task is severely hindered by substantial domain gaps and the physical intractability of capturing real-world continuous ground-truth. To this end, we propose a novel data synthesis framework to generate high-fidelity DCSZ training sequences, named ZoomGS v2. It explicitly decouples geometric trajectories and color transformations from 3D representation modeling, thereby enabling extensive data generation from easily accessible single-camera 3D scenes. Furthermore, we design ZoomFI, a specialized VFI network tailored to the DCSZ task, which leverages complementary bidirectional warping for comprehensive dual-camera information fusion. Extensive experiments on test datasets of two mobile phones demonstrate that training with our constructed DCSZ data significantly improves the performance of various VFI models. Moreover, the proposed ZoomFI achieves state-of-the-art results in both quantitative metrics and visual quality.
DCSZ Overview
(a) Existing dual-camera digital zoom. For zoom transitions between UW and W cameras (i.e., from ×0.6 to ×1.0), commercial smartphones generally crop the corresponding region from the UW image and resize it to the target dimensions. When the zoom factor transitions from ×0.9 to ×1.0, the active lens switches abruptly from UW to W, inducing severe geometric perspective and color rendition discontinuities in the preview.
(b) The proposed dual-camera smooth zoom (DCSZ) pipeline.
(c) Visual illustration of geometric and color discontinuities in conventional dual-camera zoom previews.
Video results of ZoomGS v2 and ZoomGS
(a) Pipline of ZoomGS. ZoomGS models the scene with a Base Gaussians and use CamTrans Module to learn spatial and photometric transitions between the two endpoint 3DGS, and then transform the Base Gaussians to target Gaussians of virtual cameras for rendering.
(b) Pipline of ZoomGS v2. ZoomGS v2 first calibrate distribution of camera and color parameters from small dual-camera data, and then applies them to Gaussians constructed from large single-camera data for 3D scene rendering. The producted example on the right illustrates that ZoomGS v2 generates images with better quality than ZoomGS.
Overview of ZoomFI
Previous VFI architectures struggle with the asymmetric corre- spondences and large geometric displacements between dual- camera views. To address this issue, we propose ZoomFI, a dedicated VFI model tailored to the DCSZ task, as shown above.
VFI Results
Quantitative comparisons of VFI models trained on data from ZoomGS v2 or from ZoomGS on real-world dataset. The results show that the VFI models fine-tuned on ZoomGS v2 achieve a significant performance improvement over ZoomGS on DCSZ task, and the performance of our ZoomFI surpasses other VFI models.
Comparisons of ZoomGS v2 and ZoomGS
Compare ZoomFI trained on data from ZoomGS v2 and ZoomGS.
Xiaomi Image Comparisons
Compare ZoomFI with RIFE, AMT and VFIMAMBA in Xiaomi dataset, x0.6→x1.0
Compare ZoomFI with RIFE, AMT and VFIMAMBA in Xiaomi dataset, x0.85→x1.0
Huawei Image Comparisons
Compare ZoomFI with RIFE, AMT and VFIMAMBA in Huawei dataset, x0.6→x1.0
Compare ZoomFI with RIFE, AMT and VFIMAMBA in Huawei dataset, x0.85→x1.0
Citation
If you find our work useful in your research, please consider citing:
@inproceedings{fan2026dcszv2,
title = {Learning Dual-Camera Smooth Zoom via Decoupled Geometric-Photometric Data Synthesis},
author = {Junjie Fan, Zhilu Zhang, Renlong Wu, Dazhi Zhang and Wangmeng Zuo},
journal = {arXiv preprint arXiv:2610.34854},
year = {2026},
eprint = {2610.34854},
archivePrefix= {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2610.34854}
}