As an important operation of image composition, object placement aims to predict the plausible placement (location, scale) for the inserted foreground object. Previous object placement methods can be divided into generative methods and discriminative methods, both of which cannot balance efficiency and effectiveness well. In this work, we propose a semi-generative method in the middle ground between them. In particular, we assign uniformly distributed anchors on the background. Then, we fuse foreground and background features to predict the rationality score for each anchor and predict plausible placement sets for positive anchors. Extensive experiments on the OPA dataset show that our method can strike a good balance between efficiency and effectiveness.
Figures & tables
Figure 1: The comparison among three paradigms for object placement learning. Given a pair of foreground with mask and background, the generative model (a) generates one or multiple plausible placements (center location (x,y) and scale s ) for the foreground. The discriminative model (b) predicts a rationality score map for each foreground scale. ri,s is the rationality score for the i -th center location and scale s . Our semi-generative model (c) predicts a rationality score ri for the i -th anchor and also predicts multiple plausible placements (center offset (oxi,oyi) and scale si ) for each positive anchor.
Figure 2: The illustration of our network structure. We use DFM to fuse foreground feature vector ff and background feature map Fb . For each fim in the fused feature map Fm , we use it to predict rationality score ri and concatenate it with random vector zt to produce a placement set Pi . The detailed architecture of our Dynamic Fusion Module (DFM) is shown at the top right corner. We concatenate background feature and foreground feature to produce the offset map for deformable convolution, which can dynamically aggregate multi-scale contextual information.
Figure 3: We suppose that an image is divided into 4×4 cells, corresponding to 16 anchors (red dot). In the left column, we show the foreground/background and the rationality score map. In the right column, we show one foreground placement, which is associated with the i -th anchor because the foreground center is within its neighborhood (red cell). The location of foreground center can be calculated based on the anchor location (xi,yi) and center offset (oxi,oyi) . The foreground placement (yellow bounding box) is represented by the foreground center location (xi+oxi,yi+oyi) and the foreground scale si .
Method
Plausibility
Diversity
Complexity
User Study ↑
Acc. ↑
FID ↓
LPIPS ↑
FLOPs(G)
Time(s)
TERSE [ 25 ]
0.104
0.679
46.94
0
37.25
0.040
PlaceNet [ 29 ]
0.109
0.683
36.69
0.160
4.76
0.039
GracoNet [ 32 ]
0.183
0.847
27.75
0.206
43.98
0.044
FOPA [ 19 ]
0.303
0.896
25.91
0.216
321.05
0.258
Ours
0.301
0.890
26.02
0.218
39.49
0.041
Table 1: Quantitative comparison of different methods on OPA dataset.
Figure 4: For each foreground-background pair, we show three composite images generated by GracoNet [ 32 ] , FOPA [ 19 ] , and our method.
Row
Method
Plausibility
Diversity
Acc. ↑
FID ↓
LPIPS ↑
1
w/o DC
0.871
26.85
0.212
2
w/o Lms
0.845
27.67
0.176
3
w/o Lpl
0.832
28.83
0.196
4
w/o Lcl
0.712
44.86
0.188
5
1 st stage
0.869
26.93
0.211
Table 2: Quantitative comparison of different ablated versions of our method on OPA dataset.
Figure 5: The comparison of rationality score map between FOPA [ 19 ] and our method. From left to right, we show the foreground, background, the rationality score map generated by FOPA and our method.
Figure 6: For each pair of foreground and background, we show three generated composite images for the same positive anchor with the highest rationality score. The neighborhood of each positive anchor is marked with red box on the background.
Compositional image generation studies how to combine objects, attributes, spatial relations, and visual references into coherent, controllable images.
Object placement is critical in image composition, requiring spatially and semantically coherent positioning of objects within diverse scenes. Existing approaches typically rely on hand-crafted rules or supervised learning on limited datasets, which restricts their generalization and interpretability, especially in open-world scenarios involving novel objects and scenes. In this work, we reformulate open-world object placement as a heuristic search task guided by reasoning from a Multimodal Large Language Model (MLLM). We introduce presto, a zero-shot, training-free framework that operates within an imaginary action space to iteratively refine object position and scale. Our coarse-to-fine search strategy ensures fast convergence, and we evaluate two decision-making variants: Metric-guided Selection and MLLM-as-a-judge. Experiments across multiple benchmarks show that presto~achieves state-of-the-art performance, particularly in previously unseen, open-world settings. Human studies further reveal that the MLLM-as-a-judge variant produces more perceptually coherent placements than metric-driven approaches, highlighting a gap between standard evaluation metrics and human visual judgment.
Weixuan Ding, Shang Liu, Hanyu Pei +1
Wuhan University, Wuhan Hubei Province, China · University of Louisville, Louisville KY, USA
We propose a method to learn explicit, class-conditioned spatial priors for object placement in natural scenes by distilling the implicit placement knowledge encoded in text-conditioned diffusion models. Prior work relies either on manually annotated data, which is inherently limited in scale, or on inpainting-based object-removal pipelines, whose artifacts promote shortcut learning. To address these limitations, we introduce a fully automated and scalable framework that evaluates dense object placements on high-quality real backgrounds using a diffusion-based inpainting pipeline. With this pipeline, we construct HiddenObjects, a large-scale dataset comprising 27M placement annotations, evaluated across 27k distinct scenes, with ranked bounding box insertions for different images and object categories. Experimental results show that our spatial priors outperform sparse human annotations on a downstream image editing task (3.90 vs. 2.68 VLM-Judge), and significantly surpass existing placement baselines and zero-shot Vision-Language Models for object placement. Furthermore, we distill these priors into a lightweight model for fast practical inference (230,000x faster).
Marco Schouten, Ioannis Siglidis, Serge Belongie +1
Technical University of Denmark · Pioneer Centre for AI · University of Copenhagen
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