GAN-based Domain Adaptation for Image-aware Layout Generation in Advertising Poster Design

Chenchen Xu; Min Zhou; Tiezheng Ge; Weiwei Xu

cs.LGApr 8, 2026

GAN-based Domain Adaptation for Image-aware Layout Generation in Advertising Poster Design

Authors: Chenchen Xu, Min Zhou, Tiezheng Ge, Weiwei Xu

Organizations: Anhui Normal University, Wuhu, China · Zhejiang University, Hangzhou, China · Shenzhen Bay Laboratory, Shenzhen, China · Alibaba Group, Hangzhou, China

Abstract

Layout plays a crucial role in graphic design and poster generation. Recently, the application of deep learning models for layout generation has gained significant attention. This paper focuses on using a GAN-based model conditioned on images to generate advertising poster graphic layouts, requiring a dataset of paired product images and layouts. To address this task, we introduce the Content-aware Graphic Layout Dataset (CGL-Dataset), consisting of 60,548 paired inpainted posters with annotations and 121,000 clean product images. The inpainting artifacts introduce a domain gap between the inpainted posters and clean images. To bridge this gap, we design two GAN-based models. The first model, CGL-GAN, uses Gaussian blur on the inpainted regions to generate layouts. The second model combines unsupervised domain adaptation by introducing a GAN with a pixel-level discriminator (PD), abbreviated as PDA-GAN, to generate image-aware layouts based on the visual texture of input images. The PD is connected to shallow-level feature maps and computes the GAN loss for each input-image pixel. Additionally, we propose three novel content-aware metrics to assess the model's ability to capture the intricate relationships between graphic elements and image content. Quantitative and qualitative evaluations demonstrate that PDA-GAN achieves state-of-the-art performance and generates high-quality image-aware layouts.

Figures & tables

Fig. 1: Examples of image-aware graphic layout generation for advertising posters. Our model generates graphic layouts (middle) with multiple elements conditioned on product images (left). Designers or even automatic rendering programs can use these layouts to render advertising posters (right).

Dataset	Status	Train $\mathcal{S}$	Test $\mathcal{S}$	Train $\mathcal{T}$	Test $\mathcal{T}$	Element classes	Content	Product category
NDN [ 28 ]	Private	0	500	0	0	Text, logo, image, button	Empty	Car
ICVT [ 44 ]	Private	105,862	11,762	0	166	Text, logo, underlay	Product	Clothing, electronics, cosmetics, etc.
PKU [ 33 ]	Public	9,974	0	0	905	Text, logo, underlay	Product	Clothing, electronics, cosmetics, etc.
CGL(Ours)	Public	54,546	6,002	120,000	1,000	Text, logo, underlay, embellishment	Product	Clothing, electronics, cosmetics, etc.

TABLE I: Different dataset of poster layout. The symbol

\mathcal{S}

(

\mathcal{T}

) represents the source (target) domain data.

Fig. 2: Examples of CGL-Dataset. The left part represents the six components of information contained in each sample of the dataset. In the upper right corner are examples of different product types, and in the lower right corner are examples with different layout positions.

Fig. 3: Domain gap visualization. To illustrate the domain gap, we manually design several posters based on clean product images (target data), and then inpaint the graphic elements to generate the corresponding source data. The visual content in the inpainted areas (highlighted with yellow boxes) appears distorted and blurred compared to the original content (highlighted with red boxes).

Fig. 4: The architecture of our network. Annotated posters (source domain data) must be inpainted before input to the model. The model has both reconstruction and GAN loss when training with source domain data, while only a GAN loss is used when training with target domain data. Please refer to Sec. IV for the definition of each loss term: $L_{PD}$ , $L_{PD}^{G}$ , and $L_{rec}$ . During the discriminator or generator pass, both inpainted and clean images are fed into the discriminator.

Model	$R_{com}\downarrow$	$R_{shm}\downarrow$	$R_{sub}\downarrow$	$cFID\downarrow$	$R_{ove}\downarrow$	$R_{und}\uparrow$	$R_{ali}\downarrow$	$FID\downarrow$	$R_{occ}\uparrow$
ContentGAN [ 4 ]	45.59	17.08	1.143	26.30	0.0397	0.8626	0.0071	9.62	93.4
Layoutprompter [ 32 ]	39.77	14.66	0.840	25.55	0.2251	0.6786	0.0068	8.67	99.5
CGL-GAN(Ours)	35.77	15.47	0.805	15.31	0.0233	0.9359	0.0098	5.10	99.6
PDA-GAN(Ours)	33.55	12.77	0.688	12.39	0.0290	0.9481	0.0105	4.98	99.7

TABLE II: Comparison with image-aware methods. Bold and underlined numbers denote the best and second best respectively.

\downarrow

(or

\uparrow

) means the smaller (or bigger) value, the better.

Model	$P_{e}\uparrow$	$P_{b}\uparrow$	$P^{*}_{e}\uparrow$	$P^{*}_{b}\uparrow$
ContentGAN [ 4 ]	18.89	19.36	18.25	16.58
Layoutprompter [ 32 ]	26.25	20.12	24.58	18.24
CGL-GAN	26.12	28.36	26.45	26.46
PDA-GAN	28.74	32.16	30.72	38.72

TABLE III: User study.

P_{e}

(

P_{b}

) represents the percentage of eligible-selected (best-selected) layouts. The symbol * denotes the professional group.

Fig. 5: Qualitative evaluation for different models. Layouts in each column are conditioned on the same image, while those in each row are generated by the same model. This figure provides a qualitative comparison and analysis of different models from three perspectives: background complexity of text elements, subject overlap, and product overlap attention maps.

Model	$R_{com}\downarrow$	$R_{shm}\downarrow$	$R_{sub}\downarrow$	$cFID\downarrow$	$R_{ove}\downarrow$	$R_{und}\uparrow$	$R_{ali}\downarrow$	$FID\downarrow$
LT [ 31 ]	40.92	21.08	1.310	27.24	0.0156	0.9516	0.0049	6.25
VTN [ 15 ]	41.77	22.21	1.323	30.14	0.0130	0.9698	0.0047	7.13
LDM [ 26 ]	41.20	27.91	1.792	32.36	0.0146	0.9532	0.0032	5.02
LD [ 27 ]	40.56	27.36	1.772	28.76	0.0116	0.9624	0.0028	4.28
CGL-GAN	35.77	15.47	0.805	15.31	0.0233	0.9359	0.0098	5.10
PDA-GAN	33.55	12.77	0.688	12.39	0.0290	0.9481	0.0105	4.98

TABLE IV: Comparison with image-agnostic methods.

LT

and

VTN

represent LayoutTransformer and LayoutVTN, respectively.

Model	Data I	Data II	Gaussian Blur	$R_{com}\downarrow$	$R_{shm}\downarrow$	$R_{sub}\downarrow$	$R_{ove}\downarrow$	$R_{und}\uparrow$	$R_{ali}\downarrow$	$R_{occ}\uparrow$
CGL-GAN	✓			33.85	13.88	0.766	0.0299	0.9351	0.0139	99.7
CGL-GAN	✓		✓	-	-	-	2.5826	-	-	-
CGL-GAN		✓	✓	35.77	15.47	0.805	0.0233	0.9359	0.0098	99.6
PDA-GAN	✓		✓	36.41	19.48	1.044	0.0244	0.9384	0.0091	99.7
PDA-GAN		✓	✓	30.63	18.14	0.833	0.0589	0.9302	0.0105	99.6
PDA-GAN	✓			33.55	12.77	0.688	0.0290	0.9481	0.0105	99.7

TABLE V: Comprehensive comparison between CGL-GAN and PDA-GAN. Data I and Data II contain 8,000 and 54,546 source domain samples, respectively. ✓indicates the experiment configuration. The symbol ”-” indicates that the model cannot complete the layout generation task since the generated element bounding boxes overlap with each other severely.

Fig. 6: More qualitative comparisons between CGL-GAN and PDA-GAN.

Fig. 7: Layouts generated by different models using source and target domain data. Inpainted images (source domain) and clean images (target domain) are both fed into PDA-GAN and CGL-GAN. The results produced by PDA-GAN are relatively consistent, indicating that our method achieves better feature alignment between the two domains.

Model- $W$	$R_{com}\downarrow$	$R_{shm}\downarrow$	$R_{sub}\downarrow$	$R_{ove}\downarrow$	$R_{und}\uparrow$	$R_{ali}\downarrow$
GD-6.0	-	-	-	9.0000	-	-
GD-1.0	-	-	-	8.9995	-	-
GD-0.01	-	-	-	4.7764	-	-
GD-0.001	34.41	13.78	0.749	0.0327	0.9299	0.0110
GD-0.0001	34.77	14.62	0.777	0.0345	0.9234	0.0122
GD-0.0	34.07	15.13	0.800	0.0350	0.9259	0.0108

TABLE VI: Ablation study with global discriminators.

W

refers to the weight of GD (or PD) module loss in the training process.

Patch size	$R_{com}\downarrow$	$R_{shm}\downarrow$	$R_{sub}\downarrow$	$R_{ove}\downarrow$	$R_{und}\uparrow$	$R_{ali}\downarrow$
12*8	-	-	-	0.9288	-	-
24*16	33.67	16.00	0.844	0.0438	0.9407	0.0075
44*30	34.03	13.02	0.752	0.0284	0.9377	0.0119
88*60	32.65	13.35	0.735	0.0325	0.9173	0.0094
350*240	33.55	12.77	0.688	0.0290	0.9481	0.0105

TABLE VII: Ablation study with PatchGAN-based methods. The input image height and width are 320 and 240 respectively.

Feature map	$R_{com}\downarrow$	$R_{shm}\downarrow$	$R_{sub}\downarrow$	$R_{ove}\downarrow$	$R_{und}\uparrow$	$R_{ali}\downarrow$
deep level	34.22	13.97	0.770	0.0396	0.9366	0.0118
fusion	35.36	14.54	0.817	0.0310	0.9513	0.0117
shallow level	33.55	12.77	0.688	0.0290	0.9481	0.0105

TABLE VIII: Quantitative ablation study on different level feature maps for the pixel-level discriminator.

Size	$R_{com}\downarrow$	$R_{shm}\downarrow$	$R_{sub}\downarrow$	$R_{ove}\downarrow$	$R_{und}\uparrow$	$R_{ali}\downarrow$
5	33.09	13.66	0.720	0.0349	0.8891	0.0086
7	31.69	14.73	0.743	0.0431	0.9185	0.0081
fusion	32.30	13.25	0.747	0.0333	0.9316	0.0095
3	33.55	12.77	0.688	0.0290	0.9481	0.0105

TABLE IX: Quantitative ablation study on PD with different kernel sizes. 3, 5, and 7 denote convolutional kernel sizes of 3×3, 5×5, and 7×7, respectively, while ’fusion’ indicates a hybrid design combining all three.

Layers	$R_{com}\downarrow$	$R_{shm}\downarrow$	$R_{sub}\downarrow$	$R_{ove}\downarrow$	$R_{und}\uparrow$	$R_{ali}\downarrow$
2	32.37	13.48	0.687	0.0360	0.9151	0.0110
6	32.97	15.25	0.792	0.0318	0.9040	0.0130
9	34.07	13.80	0.743	0.0279	0.9295	0.0086
3	33.55	12.77	0.688	0.0290	0.9481	0.0105

TABLE X: Quantitative ablation study on different level feature maps for the pixel-level discriminator. 2, 3, 6, and 9 denote the number of transposed convolutional layers used in the PD module.

smoothing	$R_{com}\downarrow$	$R_{shm}\downarrow$	$R_{sub}\downarrow$	$R_{ove}\downarrow$	$R_{und}\uparrow$	$R_{ali}\downarrow$
without	33.61	14.04	0.718	0.0346	0.9188	0.0106
two-side	33.66	14.67	0.794	0.0334	0.9297	0.0098
one-source	32.20	15.23	0.799	0.0431	0.9234	0.0085
one-target	33.55	12.77	0.688	0.0290	0.9481	0.0105

TABLE XI: Ablation study with different label smoothing choices. The first row is the model without label smoothing. Two-side: set 0 to 0.2 and 1 to 0.8; one-source: set 1 to 0.8; and 0ne-target: set 0 to 0.2.

Fig. 8: Impact of Gaussian blur. Layouts in each row are generated using the same input image, while those in each column correspond to different input samples. “PDA-GAN with Gaussian blur” indicates that the input data is preprocessed with Gaussian blurring. The blue-numbered boxes in the middle show enlarged views of the regions marked with yellow numbers. The left part of the vertical dotted line displays the input images, while the right part displays the corresponding product attention heatmaps.

Model	$R_{com}\downarrow$	$R_{shm}\downarrow$	$R_{sub}\downarrow$	$R_{ove}\downarrow$	$R_{und}\uparrow$	$R_{ali}\downarrow$
Ours’	34.07	15.13	0.800	0.0350	0.9259	0.0108
Ours	33.55	12.77	0.688	0.0290	0.9481	0.0105

TABLE XII: Quantitative ablation study on PD. Ours’ refers to our model of PDA-GAN without the PD module.

Backbone	$R_{com}\downarrow$	$R_{shm}\downarrow$	$R_{sub}\downarrow$	$R_{ove}\downarrow$	$R_{und}\uparrow$	$R_{ali}\downarrow$
D-DETR	32.48	13.03	0.698	0.1085	0.9423	0.0078
DETR	33.55	12.77	0.688	0.0290	0.9481	0.0105

TABLE XIII: Quantitative ablation study on different backbones. D-DETR denotes deformable DETR.

Model	$R_{com}\downarrow$	$R_{shm}\downarrow$	$R_{sub}\downarrow$	$R_{ove}\downarrow$	$R_{und}\uparrow$	$R_{ali}\downarrow$
Ours*	36.71	20.14	1.036	0.0475	0.9376	0.0068
Ours	33.55	12.77	0.688	0.0290	0.9481	0.0105

TABLE XIV: Quantitative ablation study on Gaussian blur. Ours* refers to the PDA-GAN model with Gaussian blur applied to the input image.

Fig. 9: Generalization ability of PDA-GAN on PKU-Dataset [ 33 ] .

Dataset	$R_{com}\downarrow$	$R_{shm}\downarrow$	$R_{sub}\downarrow$	$R_{ove}\downarrow$	$R_{und}\uparrow$	$R_{ali}\downarrow$
PKU [ 33 ]	33.88	18.03	0.881	0.0445	0.9411	0.0127
CGL(ours)	33.55	12.77	0.688	0.0290	0.9481	0.0105

TABLE XV: Comparison of PDA-GAN trained on different datasets.

Fig. 10: Language-guided layout generation with varying element counts and types. Each row corresponds to the same background image. The left part shows layouts generated from prompts specifying the number of elements (e.g., “generate three elements”, 2 to 6 from left to right), while the right part shows layouts guided by prompts specifying element types (e.g., “generate text and underlay”).

Fig. 11: Failure cases. The first to third rows show layouts generated by models with the global discriminator (GD), patch discriminator (PatchD), and our PD module, respectively.

Generative Adversarial Networks

Generative Adversarial Networks (GANs) train competing generators and discriminators to synthesize data, translate images, and model complex distributions.

See all 177 papers

iPoster: Content-Aware Layout Generation for Interactive Poster Design via Graph-Enhanced Diffusion Models
Mar 31, 2026Xudong Zhou, Jinyuan Liang, Qiuyi Guo +1Constrained Generative Modeling Diffusion Models
Beijing Institute of Technology, Beijing, China
UniLayDiff: A Unified Diffusion Transformer for Content-Aware Layout Generation
Dec 9, 2025Zeyang Liu, Le Wang, Sanping Zhou +4Diffusion Transformer Conditional Generative Modeling
College of Artificial Intelligence, Xi’an Jiaotong University · Amazon.com, Inc. · Pixocial Technology
Creating Power Distribution Network Layouts Using Generative Adversarial Networks and Image-Based Representations
Jul 7, 2026Juan Manuel Garcia-Perez, Carlos MateoSynthetic Data Generation Conditional Generative Modeling
School of Engineering (ICAI), Comillas Pontifical University, Spain · Institute for Research in Technology (IIT), School of Engineering (ICAI), Comillas Pontifical University, Spain

Mar 31, 2026J/K move · Enter open · S save

iPoster: Content-Aware Layout Generation for Interactive Poster Design via Graph-Enhanced Diffusion Models

Xudong Zhou, Jinyuan Liang, Qiuyi Guo, Guozheng Li

Beijing Institute of Technology, Beijing, China

Constrained Generative Modeling Diffusion Models

Open paper AskPDF arXiv

We present iPoster, an interactive layout generation framework that empowers users to guide content-aware poster layout design by specifying flexible constraints. iPoster enables users to specify partial intentions within the intention module, such as element categories, sizes, positions, or coarse initial drafts. Then, the generation module instantly generates refined, context-sensitive layouts that faithfully respect these constraints. iPoster employs a unified graph-enhanced diffusion architecture that supports various design tasks under user-specified constraints. These constraints are enforced through masking strategies that precisely preserve user input at every denoising step. A cross content-aware attention module aligns generated elements with salient regions of the canvas, ensuring visual coherence. Extensive experiments show that iPoster not only achieves state-of-the-art layout quality, but offers a responsive and controllable framework for poster layout design with constraints.

Abstract

Generative Adversarial Networks

Explore similar work

iPoster: Content-Aware Layout Generation for Interactive Poster Design via Graph-Enhanced Diffusion Models

UniLayDiff: A Unified Diffusion Transformer for Content-Aware Layout Generation

Creating Power Distribution Network Layouts Using Generative Adversarial Networks and Image-Based Representations