We present Mover360, a controllable object manipulation framework for 360° images. Unlike perspective images, 360° images in equirectangular projection (ERP) exhibit horizontal wrap-around, latitude-dependent distortion, and global scene continuity, which makes object-level edits difficult for existing perspective editors to produce and for users to specify. To address this, Mover360 centers on object Translation (relocating a specified object within an existing panorama) while supporting reference-guided Insert and Remove as auxiliary tasks. Its interface unifies point-, bbox-, and mask-guided control by encoding each task into a fixed prompt and a compact, ERP-aligned instruction map. In the default point mode, a single click relocates an object, allowing the model to infer a plausible size, support, and illumination using panoramic context and an auxiliary depth condition. Structurally, Mover360 is a lightweight adaptation of a pretrained diffusion transformer. To generate paired supervision, we construct a UE5 data-generation pipeline with surface-aware object placement and randomized illumination, yielding large-scale paired data and a dual-domain benchmark of synthetic and real panoramas with ground truth for all three tasks. Across both test domains and two evaluation protocols, Mover360 outperforms strong baselines for perspective editing, insertion, and inpainting in reconstruction fidelity, semantic consistency, and distributional quality. Code and our benchmark dataset are available at https://zhonghaoyi.github.io/Mover360/.
Figures & tables
Figure 1: Anatomy of one Mover360 editing tuple from the UE5 test set (top) and from the real captured test set (bottom). Each tuple contains a source-state panorama, a target-state panorama, an object-absent background panorama, and a perspective reference image, and therefore supervises all three tasks: Translation (source → target), Remove (source → background), and Insert (background + reference → object present). The left column overlays all three guidance granularities of the instruction map: fine object masks (contours), bbox regions (translucent rectangles), and the Gaussian target point placed at the bottom-center of the target box (blue dot), with the source in red and the target in green. Note that the source object of the real example straddles the horizontal wrap-around boundary of the ERP.
Figure 2: Examples from the Mover360 data generation pipeline. The top two rows show randomized nearby lamps, including variations in intensity, temperature, scattering or beam angle, and position. In the third row, the first two examples show outdoor/global light-source variation, and the last example shows one object path captured by three panoramic cameras.
Train (UE5)
Test (UE5)
Test (real)
Editing samples
12,150 seq.
210
50
Sampled pairs per epoch
60,750
–
–
Unique objects
608
48
15
Ground truth for
Translation / Remove / Insert
Object size (% of sphere)
0.69 [0.27, 1.72]
1.04 [0.40, 2.15]
0.79 [0.49, 1.72]
Translation ( ∘ )
23.4 [11.9, 42.3]
31.4 [18.0, 47.5]
50.9 [34.2, 103.0]
Table 1: Statistics of the Mover360 training corpus and the dual-domain test benchmark. Distribution rows report median [interquartile range], computed from the ERP object masks: object size is the solid-angle fraction of the sphere, translation is the wrap-aware great-circle displacement of the mask centroid (sampled training pairs for the corpus, source-to-target pairs for the test sets), and elevation is the centroid latitude.
Figure 3: Distributions of sphere-aware statistics for the Mover360 training corpus and the two test sets: object size as a percentage of the viewing sphere (log scale), great-circle translation distance, and object elevation. The two synthetic sets share similar distributions with a wider size range, while the real test set deliberately stresses long-range translations (median 50.9∘ , with a quarter of the moves beyond 103∘ ).
Figure 4: Overview of Mover360. The framework takes an input ERP panorama, a fixed task prompt, an ERP-aligned instruction map, an auxiliary DA 2 depth map, and an optional reference image for insertion. Condition latents are concatenated and patchified before being processed by the Flux2-Klein diffusion transformer. We adapt the native 4D RoPE coordinates to distinguish text tokens, noisy target panorama tokens, and multiple conditioning image slots. LoRA adapters are trained while the reference branch is disabled for translation and remove. VAE encoding and decoding are wrapped with horizontal circular pad-and-unpad to better respect panorama periodicity.
Figure 5: Qualitative comparison for the Translation task. Baselines often leave residual content at the source location, distort the moved object, or produce less plausible contact with the target surface. Mover360 uses the point or region instruction to move the object while preserving the surrounding panorama.
Figure 6: Qualitative comparison for the Insert task. Reference-guided baselines may copy viewpoint-specific appearance or produce inconsistent object scale. Mover360 better adapts the inserted object to the target region and the panorama context.
Figure 7: Qualitative comparison for the Remove task. LaMa is a strong specialized inpainting baseline, but Mover360 performs removal within the same unified framework used for translation and insertion.
Figure 8: Qualitative comparison under the perspective protocol . Rows show Translation (UE5), Insert (real), and Remove (UE5); columns show the input panorama with the instruction regions overlaid (source in red, target in green, reference inset for Insert), the square perspective view given to the baseline with the provided target box, the baseline’s edited view, the baseline’s final panorama after re-projection, Mover360 (mask), and the ground truth. For Translation, the baseline crops the source-removed panorama. Mask-conditioned baselines struggle with global consistency due to local cropping: the relocated car and inserted backpack lack the cast shadows and reflected illumination seen in Mover360 and ground truth. For Remove, the baseline fails to eliminate the long shadow extending beyond its perspective crop, whereas Mover360 removes the entire shadow.
UE5 test set (210)
Real test set (50)
Task
Method
FAED ↓
PSNR ↑
SSIM ↑
LPIPS ↓
DINOv3 ↑
FID ↓
FAED ↓
PSNR ↑
SSIM ↑
LPIPS ↓
DINOv3 ↑
FID ↓
Translation
DragAnything [ 29 ]
8.797
12.68
0.351
0.607
0.288
208.0
24.362
10.86
0.422
0.679
0.215
240.0
AnyDoor [ 4 ]
1.157
22.18
0.571
0.352
0.687
116.0
4.167
22.00
0.673
0.345
0.611
167.0
Paint-by-Example [ 33 ]
1.504
21.75
0.564
0.337
0.437
133.0
1.888
22.29
0.626
0.352
0.454
199.0
OmniPaint [ 37 ]
1.196
22.02
0.696
0.318
0.662
117.0
2.340
21.43
0.662
0.323
0.656
167.0
Insert-Anything [ 19 ]
0.344
24.23
0.809
0.297
0.739
108.9
0.229
26.23
0.802
0.279
0.718
151.0
Table 2: Quantitative comparison under the ERP protocol , where baselines operate directly on equirectangular panoramas, reported separately on the UE5 (210 samples per task) and real (50 samples per task) test sets. FAED and PSNR are computed on full ERP panoramas; SSIM, LPIPS, DINOv3 similarity, and FID are computed on perspective views centered on the edit region. Mover360 (point) uses only a target point. All baselines receive the known-size target bbox or mask.
UE5 test set (210)
Real test set (50)
Task
Method
FAED ↓
PSNR ↑
SSIM ↑
LPIPS ↓
DINOv3 ↑
FID ↓
FAED ↓
PSNR ↑
SSIM ↑
LPIPS ↓
DINOv3 ↑
FID ↓
Translation
AnyDoor [ 4 ]
0.451
23.39
0.798
0.291
0.727
103.7
0.307
26.56
0.767
0.277
0.675
140.6
Paint-by-Example [ 33 ]
0.559
23.30
0.817
0.302
0.513
131.6
0.459
25.42
0.768
0.303
0.561
176.4
Insert-Anything [ 19 ]
0.323
24.52
0.822
0.256
0.769
97.8
0.214
27.88
0.784
0.250
0.727
132.4
OmniPaint [ 37 ]
0.461
23.36
0.813
0.285
0.692
107.4
0.245
27.17
0.784
0.268
0.653
135.6
Mover360 (point)
0.265
24.25
0.830
0.270
0.721
97.8
0.294
26.04
0.781
0.274
0.669
144.5
Table 3: Quantitative comparison under the perspective protocol , where each baseline edits a locally undistorted, adaptively framed square perspective view (FOV clipped to [35∘,150∘] ) and only the edited region is re-projected into the panorama. This adaptation is strictly favorable to the baselines. Mover360 operates natively on the ERP; for fairness, its output pixels outside the perspective-view crop are likewise replaced by the input panorama, so its FAED and PSNR differ from Table 2 while the perspective-view metrics are unchanged. DragAnything is omitted because its video-based dragging does not transfer to this protocol.
UE5 test set (210)
Real test set (50)
Task
Method
FAED ↓
PSNR ↑
SSIM ↑
LPIPS ↓
DINOv3 ↑
FID ↓
FAED ↓
PSNR ↑
SSIM ↑
LPIPS ↓
DINOv3 ↑
FID ↓
Translation
Mover360 w/o depth
0.351
24.42
0.813
0.275
0.708
100.0
0.374
26.53
0.752
0.297
0.655
148.5
Mover360 (point)
0.336
24.14
0.830
0.270
0.721
97.8
0.357
26.82
0.781
0.274
0.669
144.5
Mover360 (bbox)
0.314
25.08
0.836
0.255
0.747
95.5
0.356
26.85
0.823
0.249
0.704
131.0
Mover360 (mask)
0.250
25.43
0.868
0.197
0.780
82.9
0.186
29.35
0.841
0.181
0.717
121.6
Insert
Mover360 w/o depth
0.438
23.42
0.821
0.355
0.671
112.0
0.493
25.59
0.734
0.335
0.625
174.1
Table 4: Ablation study on the auxiliary depth condition and the spatial-instruction granularity (point/bbox/mask), reported separately on the UE5 and real test sets. FAED and PSNR are computed on full ERP panoramas; the remaining metrics are computed on perspective views centered on the edit region. The w/o-depth variant uses point guidance for Translation and Insert and bbox guidance for Remove.
Figure 9: Failure cases of Mover360. Top: translating a large-area light source. The emitter is relocated to the target point, but the wide illumination it casts is not re-synthesized: the brightly lit region anchored to the original position survives (dashed boxes), decoupling the moved source from the scene lighting; compact emitters do not trigger this failure. Bottom: translating a building-scale structure. The building is relocated, but additional structures absent from the input are hallucinated near the target (dashed box).
Figure S1 : Additional Translation results. Each pair shows the input panorama with the source region, the motion arrow, and the Gaussian target point overlaid, and the edited panorama with a magnified view of the target region (dashed box).
Figure S2 : Additional Insert results. Each pair shows the input panorama with the reference object (inset) and the target point overlaid, and the edited panorama with a magnified view of the inserted object (dashed box).
Figure S3 : Additional Remove results. Each pair shows the input panorama with the source object marked, and the edited panorama with a magnified view of the completed background (dashed box).
Figure S4 : Qualitative ablation of the auxiliary depth condition under the default point guidance. Asked to relocate the object to the target point (arrow), the w/o-depth model renders it at an implausibly large scale, whereas the depth-conditioned model infers a projected size consistent with the scene depth at the target location (insets).
Interactive panoramic video generation aims to synthesize immersive 360\textdegree{} videos that remain visually coherent while following user-specified camera trajectories during exploration. However, progress is limited by a coupled data-and-model gap: existing panoramic video datasets are often short, weakly annotated, or lack camera trajectories, while existing camera-controlled video generation models are designed for perspective videos and do not directly support panoramic geometry. In this paper, we introduce MUGEN and Wan360 to address these limitations. MUGEN is a large-scale real-world panoramic video dataset tailored to interactive 360-degree world exploration, comprising over 1,300 hours of at least 4K panoramic videos with rich semantic and geometric annotations. Built on MUGEN, we further present Wan360, a camera-controllable interactive panoramic video generation model. Panoramic videos are commonly represented by EquiRectangular Projection (ERP), which unfolds a spherical 360-degree view into a rectangular frame with cyclic longitude seams and pole distortions. To this end, Wan360 introduces three parameter-free ERP-aware components: periodic longitude RoPE for seam-consistent positional encoding, ERP-aware padding for reducing boundary artifacts, and random roll yaw for consistent learning. For camera control, Wan360 uses a panoramic Plücker embedding that represents camera motion with ERP rays rather than perspective pinhole rays. Experiments show that MUGEN serves as a data foundation for panoramic world exploration, and that Wan360 enables high-quality, temporally coherent, camera-controllable 360-degree video generation.
Jiaming Tan, Zhen Li, Shuwei Shi +6
Alaya Lab · Beijing Institute of Technology · Shanghai Innovation Institute
Diffusion models provide powerful priors for 2D image completion, but these priors are learned on bounded planar images and do not transfer directly to 360∘ panoramas. Perspective observations and spherical panoramas differ in both projective geometry and topology: viewpoint-dependent distortion complicates spatial correspondence, while Equirectangular Projection (ERP) panoramas exhibit intrinsic S1 periodicity that standard Euclidean architectures do not preserve. We present Gimbal360, a unified framework that adapts planar diffusion priors to spherical panoramic completion by standardizing these geometric and topological structures. Our Canonical Viewing Space expresses projective distortion as a fixed function of latitude, providing a consistent interface between perspective inputs and spherical panoramas. To map unposed in-the-wild images into this space, Differentiable Projective Canonicalization projects a dense correspondence field onto a 3-DoF rigid projection manifold without requiring camera parameters at inference. We further introduce Topologically Equivariant Generation, which enforces latent shift equivariance to preserve continuity across the periodic ERP boundary. Together, these designs allow diffusion to operate in a representation whose geometry and topology are explicitly aligned with the spherical domain. We also introduce Horizon360, a curated large-scale dataset of gravity-aligned panoramic environments. Extensive experiments show that Gimbal360 achieves state-of-the-art visual fidelity and seam continuity in 360∘ scene completion.
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
Or Hirschorn, Aaron Olender, Eli Alshan +3
1Amazon Prime Video · 2Tel-Aviv University · 3Hebrew University of Jerusalem