Realistic image composition makes a pasted object appear naturally present in a scene while preserving the pasted layout and fine foreground and background details. Reference-based insertion methods preserve overall object identity but can alter pose and fine details, often at substantial computational cost. Foreground harmonization and shadow generation address individual aspects of this task, but neither alone ensures natural integration. An efficient, unified solution remains an open challenge, and large-scale, high-quality datasets and benchmarks for this task are lacking. To address these challenges, we introduce Paste2Place, an end-to-end, physics-guided flow-based model that learns a direct latent bridge from pasted inputs to realistic composites, guided by image-space corrections analytically derived from predicted physical components of object and scene appearance changes. To support training and evaluation, we introduce P2P-Data with 86K manually reviewed synthetic and real-scene training pairs, including dense physical annotations from our Blender rendering and validation pipeline for the synthetic subset. P2P-Bench combines paired compositions and manually pasted inputs from collected images into 1.5K examples for evaluating composition under user-chosen placements. Extensive qualitative and quantitative experiments demonstrate that Paste2Place achieves substantial advantages in image composition quality and inference speed across diverse scenarios.
Figures & tables
Figure 1: Different objectives in object placement. Visual comparison of three related but distinct tasks: image composition, reference-based insertion, and foreground relighting.
Figure 2: Construction of P2P-Data and P2P-Bench. (a) Controlled Blender rendering provides aligned composition pairs and physical annotations. (b) OBER-derived pairs match appearance-perturbed inputs to the original photographs. (c) Manually pasted foregrounds in new scenes form the unpaired Wild benchmark.
Figure 3: Overview of Paste2Place. (a) A source-anchored latent bridge maps the pasted input to the composite in one step. (b) Early DiT context and image–mask features support physical component prediction. (c) Analytic recomposition yields object and environment corrections that guide later DiT blocks within the same forward pass.
Figure 4: Predicted physical corrections. (a) Object and environment corrections versus rendered references. (b) Local crops and profiles along dashed lines. Maps and profiles share an asinh(ΔY/s) scale for linear-RGB luminance changes ΔY . Map colors indicate attenuation (blue) and increase (orange).
Figure 5: Qualitative comparison on P2P-Bench. The examples highlight differences in object relighting, scene effects, and fine-detail preservation. Green boxes mark the enlarged bottle-label regions.
Metric
MV-CoLight
OmniPaint
PS-Diffusion
DreamFuse
AnyDoor
DIRECT
SHINE
Insert Anything
Qwen-Image
Seedream
Nano Banana
Ours
PSNR ↑
24.35
21.67
21.27
23.70
20.43
22.01
20.24
21.84
22.32
22.11
23.03
30.83
SSIM ↑
0.9376
0.8672
0.7830
0.9358
0.8262
0.9344
0.7900
0.9009
0.9261
0.9212
0.9290
0.9427
LPIPS ↓
0.0686
0.1043
0.2108
0.0665
0.1618
0.0687
0.1859
0.0976
0.0678
0.0725
0.0687
0.0396
fMSE ↓
3091.1
5744.9
3986.7
3640.7
6594.0
6700.9
6543.0
6026.6
5853.4
5773.6
4041.5
713.2
Effect MAE ↓
0.1343
0.1344
0.1414
0.1571
0.1676
0.1687
0.1709
0.1722
0.1433
0.1504
0.1374
0.0589
Latency (s) ↓
1.289
20.415
16.795
41.518
8.500
70.257
115.424
22.661
15.677
65.619
27.559
0.743
Table 1: Quantitative comparison on P2P-Bench. Bold and underlined values indicate the best and second-best reported results.
Figure 6: Human ratings of composition quality. Methods are ordered by mean score; violin widths show rating density.
Setting
PSNR ↑
SSIM ↑
LPIPS ↓
fMSE ↓
Effect MAE ↓
FLUX.2 9B (fine-tuned)
24.33
0.8645
0.0742
1649.3
0.0980
w/o physical branch
28.58
0.9265
0.0493
841.9
0.0725
Spatially shifted guidance
28.20
0.9298
0.0584
905.6
0.0801
Ours
30.83
0.9427
0.0396
713.2
0.0589
Table 2: Bridge adaptation and physical guidance. Results on 1,250 paired examples compare FLUX.2 9B fine-tuned on P2P-Data with Paste2Place variants.
Compositional image generation studies how to combine objects, attributes, spatial relations, and visual references into coherent, controllable images.
Despite strong single-turn performance, diffusion-based image compositing often struggles to preserve coherent spatial relations in pairwise or sequential edits, where subsequent insertions may overwrite previously generated content and disrupt physical consistency. We introduce PICS, a self-supervised composition-by-decomposition paradigm that composes objects in parallel while explicitly modeling the compositional interactions among (fully-/partially-)visible objects and background. At its core, an Interaction Transformer employs mask-guided Mixture-of-Experts to route background, exclusive, and overlap regions to dedicated experts, with an adaptive α-blending strategy that infers a compatibility-aware fusion of overlapping objects while preserving boundary fidelity. To further enhance robustness to geometric variations, we incorporate geometry-aware augmentations covering both out-of-plane and in-plane pose changes of objects. Our method delivers superior pairwise compositing quality and substantially improved stability, with extensive evaluations across virtual try-on, indoor, and street scene settings showing consistent gains over state-of-the-art baselines. Code and data are available at https://github.com/RyanHangZhou/PICS
Generative image composition aims to regenerate the given foreground object in the background image to produce a realistic composite image. Some high-authenticity methods can adjust foreground pose/view to be compatible with background, while some high-fidelity methods can preserve the foreground details accurately. However, existing methods can hardly achieve both goals at the same time. In this work, we propose a two-stage strategy to achieve both goals. In the first stage, we use high-authenticity method to generate reasonable foreground shape, serving as the condition of high-fidelity method in the second stage. The experiments on MureCOM dataset verify the effectiveness of our two-stage strategy. The code and model have been released at https://github.com/bcmi/OSInsert-Image-Composition.
Jingyuan Wang, Li Niu
MoE Key Lab of Artificial Intelligence, Shanghai Jiao Tong University
3D Gaussian Splatting (3DGS) is widely used to capture and render real scenes. Compositing objects from one capture into another has applications in many domains, such as VFX, architecture and interior design, or marketing. However, extracting an object from a source scene and naively pasting it into a target scene will fail to produce realistic results due to the different lighting conditions between the two scenes. To address this problem, we introduce a diffusion model that harmonizes naively composited images with inconsistent lighting. The model is trained with a heterogeneous dataset of image pairs (inconsistent composite input, consistent output), combining synthetic, generated, and real data. Our complete 3D solution allows a user to extract an object from the source scene and composite it into the target scene. From this, the (inconsistent) views of the target scene with the composite object are rendered. Our diffusion model harmonizes each one of these views, which are finally consolidated in a 3DGS representation with a post-optimization step. Our method provides visually compelling results, making object transfer between 3DGS easy to use and significantly improving quality compared to previous methods.
Nicolás Violante, George Kopanas, Linus Franke +2
Inria, Université Côte d’Azur, France · Google DeepMind, United States · Eyeline Labs, United Kingdom