ReX-Shot: Single-Image Rephotography via Geometry- and Camera-Grounded Generation
Authors: Ruiqi Zhang, Hao Zhu, Wenhao Zhang, Qi Zhang, Junqi Shi, Ming Lu, Xun Cao, Zhan Ma
Organizations: School of Electronic Science and Engineering, Nanjing University, Nanjing, 210023, China · vivo Mobile Communication Co., Ltd., Xi’an, 710055, China
Single-image rephotography aims to synthesize new shots of a scene from a single reference image with specified viewpoints, focal lengths, and photographic effects, which are intrinsically coupled in imaging. Existing methods typically treat these factors separately and struggle under joint control: novel-view synthesis may introduce geometric distortions under focal-length changes, while super-resolution and instruction-guided editing remain confined to 2D and cannot reliably extend detail restoration or appearance control to novel viewpoints. We attribute these limitations to imperfect single-image 3D reconstruction and the sampling limit of continuous focal-length enlargement. To reduce projection bias from geometric errors, we use implicitly transformed foundation-model features for robust target-view guidance. We further formulate focal-length enlargement as a geometry-guided super-resolution problem and exploit generative detail priors to recover details lost during sparse 3D resampling. Built on this 3D-aware generative backbone, we lift photographic-effect control from 2D filtering to 3D-aware appearance editing, preserving content consistency across viewpoints and focal lengths. These components form ReX-Shot, a geometry- and camera-grounded generative framework for single-image rephotography. To our knowledge, ReX-Shot is the first unified framework to jointly control viewpoint, focal length, and parameterized photographic effects from a single image. Experiments show that ReX-Shot outperforms representative baselines across all three controls while enabling near-real-time interactive rephotography.
Figures & tables
Fig. 1: Overview of ReX-Shot for single-image rephotography. Left: Multidimensional rephotography capabilities. ReX-Shot virtually recaptures the reference scene as target shots with controllable camera geometry and photographic effects, including viewpoint, focal length, color temperature, exposure, saturation, and contrast. Right: Quantitative performance comparison. ReX-Shot supports viewpoint, focal-length, and photographic-effect control while outperforming representative baselines from different method categories. All results in the radar chart are evaluated on the DL3DV-140 [ 1 ] test set. For each metric, we globally normalize the scores across the three tasks, and the numerical labels on the concentric grid rings shown in the Viewpoint section also apply to the same metric in the Focal Length and Photographic Effects sections. The thick solid, thin solid, dashed, and dotted lines denote ReX-Shot, Novel View Synthesis, Super Resolution, and Instruction-Guided Editing, respectively.
Fig. 2: Pipeline of ReX-Shot. Given a single reference image, ReX-Shot first estimates a per-pixel point cloud and camera parameters, and then synthesizes rephotographed images with controllable viewpoint, focal length, and photographic effects. The pipeline comprises two stages. 1. Dual-space geometry and camera grounding combines pixel-space projection, which renders the estimated point cloud under the target camera, with feature-space projection, which implicitly transforms VGGT [ 29 ] and DINOv2 [ 31 ] features into camera-aware tokens through PRoPE-based cross-attention [ 41 ] . 2. Generative rephotography leverages the detail prior of a one-step generative super-resolution model to recover details missing from sparse 3D resampling under the guidance of dual-space conditions, while a parameterized photographic-effect encoder injects continuous effect controls into the DiT blocks. Snowflakes and flames denote frozen and trainable modules, respectively, and ⊕ denotes concatenation.
Fig. 3: Qualitative comparison with novel-view-synthesis baselines under viewpoint and focal-length control. For each example, the first row shows results at the target viewpoint with the focal-length scale fixed at 1.0× . The second and third rows keep the target viewpoint fixed and enlarge the focal length by 2.0× and 3.0× , respectively. The input reference and ground-truth target images are shown in the leftmost and rightmost columns, respectively.
Method
DL3DV
MipNeRF360
Tanks & Temples
Average
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
AnySplat
16.51
0.464
0.461
14.66
0.457
0.477
13.61
0.425
0.521
14.93
0.449
0.486
CameraCtrl
14.45
0.391
0.480
13.71
0.400
0.604
12.69
0.411
0.545
13.62
0.401
0.543
LVSM
18.87
0.526
0.383
15.48
0.498
0.636
15.62
0.514
0.511
16.66
0.513
0.510
ViewCrafter
15.84
0.417
0.424
16.13
0.443
0.453
14.18
0.433
0.487
15.38
0.431
0.455
PE-Field
16.34
0.429
0.427
15.95
0.443
0.486
14.31
0.433
0.485
15.53
0.435
0.466
TABLE I: Quantitative comparison under joint viewpoint and focal-length control.
Method
DL3DV
MipNeRF360
Tanks & Temples
Average
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
AnySplat
18.07
0.638
0.223
12.39
0.479
0.419
12.65
0.526
0.386
14.37
0.548
0.343
CameraCtrl
17.63
0.481
0.273
15.56
0.438
0.503
14.53
0.469
0.428
15.91
0.463
0.401
LVSM
19.82
0.564
0.210
15.74
0.459
0.523
15.86
0.510
0.388
17.14
0.511
0.374
ViewCrafter
19.94
0.570
0.205
17.37
0.462
0.380
15.30
0.470
0.397
17.54
0.501
0.327
PE-Field
19.55
0.510
0.236
18.59
0.491
0.344
16.49
0.495
0.348
18.21
0.499
0.309
TABLE II: Quantitative comparison under viewpoint-only control.
Method
DL3DV
MipNeRF360
Tanks & Temples
Average
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
AnySplat
20.21
0.676
0.398
25.39
0.797
0.285
20.90
0.695
0.365
22.17
0.723
0.349
CameraCtrl
15.65
0.430
0.409
15.20
0.426
0.488
14.50
0.448
0.433
15.12
0.435
0.443
LVSM
22.16
0.628
0.259
24.37
0.689
0.279
22.98
0.681
0.247
23.17
0.666
0.262
ViewCrafter
16.40
0.428
0.427
19.70
0.551
0.403
15.90
0.486
0.408
17.33
0.488
0.413
PE-Field
17.08
0.453
0.390
18.47
0.509
0.359
16.37
0.489
0.389
17.31
0.484
0.379
TABLE III: Quantitative comparison under focal-length-only control.
Fig. 4: Qualitative focal-length-only comparison between TSD-SR and ReX-Shot under continuous focal-length enlargement.
Metric
Effect
ICEdit
IP2P
ReX-Shot
PSNR ↑
Color Temp.
23.28
20.23
31.72
Exposure
17.25
16.13
30.18
Saturation
19.92
27.75
30.13
Contrast
19.86
20.64
28.59
Average
20.08
21.19
30.16
SSIM ↑
Color Temp.
0.909
0.781
0.913
TABLE IV: Quantitative comparison under photographic-effect control.
Fig. 5: Qualitative comparison with instruction-guided image editing baselines under color-temperature control.
Fig. 6: Qualitative comparison with instruction-guided image editing baselines under exposure control.
Fig. 7: Qualitative comparison with instruction-guided image editing baselines under saturation control.
Fig. 8: Qualitative comparison with instruction-guided image editing baselines under contrast control.
Fig. 9: Inference latency comparison between ReX-Shot and representative feed-forward and generative methods. When applicable, end-to-end latency is decomposed into reconstruction, rendering for camera-geometry control, and effect editing for photographic-effect control.
Method
Focal Length Only
Viewpoint Only
Joint Control
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
SD3-FT
25.42
0.764
0.141
20.82
0.590
0.216
20.69
0.566
0.230
TSD-SR-FT
29.64 (+4.22)
0.867 (+0.103)
0.061 (−0.080)
22.30 (+1.48)
0.658 (+0.068)
0.134 (−0.082)
21.85 (+1.16)
0.620 (+0.054)
0.153 (−0.077)
TSD-SR-FT+FP
30.00 (+0.36)
0.878 (+0.011)
0.058 (−0.003)
23.13 (+0.83)
0.703 (+0.045)
0.122 (−0.012)
23.05 (+1.20)
0.679 (+0.059)
0.137 (−0.016)
TABLE V: Quantitative comparison across SD3-FT, TSD-SR-FT, and TSD-SR-FT+FP. Parentheses report changes relative to the preceding row.
Fig. 10: Qualitative results under joint control of viewpoint, focal length, and color temperature.
Fig. 11: Qualitative comparison across SD3-FT, TSD-SR-FT, and TSD-SR-FT+FP. The top row shows the generated images, and the bottom row shows the corresponding error maps. All error maps are visualized using the same color scale, with warmer colors indicating larger errors.
Fig. 12: PCA visualization of the projected output features, with the first three principal components mapped to RGB. The two rows correspond to two target viewpoints, and the columns show focal-length scales of 1.0× , 2.0× , and 3.0× .
Single-image refocusing changes the focal plane and depth of field of a photograph after capture. It is inherently interactive: users try many focal points and blur strengths on one image, so each edit should be fast and change only the blur. Recent generative methods, however, rerun large diffusion models for every edit. We observe that the two sub-problems differ in kind: recovering the all-in-focus image is ill-posed and needs a generative prior, whereas rendering blur is largely determined by depth and the focal plane. We therefore propose FlashReFocus, which runs diffusion once per photograph and renders every edit with a lightweight network. To address the ill-posed restoration, a one-step diffusion deblurrer with a decoder refiner recovers fine detail efficiently while balancing fidelity and perceptual quality. For precise control, the renderer is conditioned on an explicit defocus map and uses defocus-bias attention, which lets pixels at similar defocus share information without interference from pixels at other depths; in-focus content stays unchanged, and editing the map yields tilt-shift and two focal planes. To further test generalization to unseen cameras, we collect 3CReal, a three-camera benchmark of 89 pairs. FlashReFocus achieves the best perceptual fidelity for zero-shot bokeh rendering on EBB! and 3CReal, refocuses 148.3x faster than GenRefocus, and takes about half a second per further edit, 462x faster.
Precise camera control for reshooting dynamic videos is bottlenecked by the severe scarcity of paired multi-view data for non-rigid scenes. We overcome this limitation with a highly scalable self-supervised framework capable of leveraging internet-scale monocular videos. Our core contribution is the generation of pseudo multi-view training triplets, consisting of a source video, a geometric anchor, and a target video. We achieve this by extracting distinct smooth random-walk crop trajectories from a single input video to serve as the source and target views. The anchor is synthetically generated by forward-warping the first frame of the source with a dense tracking field, which effectively simulates the distorted point-cloud inputs expected at inference. Because our independent cropping strategy introduces spatial misalignment and artificial occlusions, the model cannot simply copy information from the current source frame. Instead, it is forced to implicitly learn 4D spatiotemporal structures by actively routing and re-projecting missing high-fidelity textures across distinct times and viewpoints from the source video to reconstruct the target. At inference, our minimally adapted diffusion transformer utilizes a 4D point-cloud derived anchor to achieve state-of-the-art temporal consistency, robust camera control, and high-fidelity novel view synthesis on complex dynamic scenes.
Creating flexible 3D scenes from a single image is vital when direct 3D data acquisition is costly or impractical. We introduce NavCrafter, a novel framework that explores 3D scenes from a single image by synthesizing novel-view video sequences with camera controllability and temporal-spatial consistency. NavCrafter leverages video diffusion models to capture rich 3D priors and adopts a geometry-aware expansion strategy to progressively extend scene coverage. To enable controllable multi-view synthesis, we introduce a multi-stage camera control mechanism that conditions diffusion models with diverse trajectories via dual-branch camera injection and attention modulation. We further propose a collision-aware camera trajectory planner and an enhanced 3D Gaussian Splatting (3DGS) pipeline with depth-aligned supervision, structural regularization and refinement. Extensive experiments demonstrate that NavCrafter achieves state-of-the-art novel-view synthesis under large viewpoint shifts and substantially improves 3D reconstruction fidelity.
Hongbo Duan, Peiyu Zhuang, Yi Liu +5
Center for Artificial Intelligence and Robotics, Shenzhen International Graduate School, Tsinghua University, Shenzhen 518055, China · Peng Cheng Laboratory, 518108, China · School of Cyber Science and Technology, Shenzhen Campus of Sun Yat-sen University, China +1