CARVIEW

MOTORHOMES

Select Language

HTTP/2 301 server: GitHub.com content-type: text/html location: https://brucew91.github.io/SEGA.github.io/ x-github-request-id: 9B8D:1F53DD:8B02AB:9C1C76:69525F76 accept-ranges: bytes age: 0 date: Mon, 29 Dec 2025 11:01:10 GMT via: 1.1 varnish x-served-by: cache-bom-vanm7210037-BOM x-cache: MISS x-cache-hits: 0 x-timer: S1767006070.487504,VS0,VE203 vary: Accept-Encoding x-fastly-request-id: 7a8b91abd3045d3dba758a9295a680a880ce5091 content-length: 162 HTTP/2 200 server: GitHub.com content-type: text/html; charset=utf-8 last-modified: Mon, 20 Oct 2025 02:25:46 GMT access-control-allow-origin: * strict-transport-security: max-age=31556952 etag: W/"68f59daa-8900" expires: Mon, 29 Dec 2025 11:11:10 GMT cache-control: max-age=600 content-encoding: gzip x-proxy-cache: MISS x-github-request-id: C93A:292AC1:8B0190:9C1B22:69525F76 accept-ranges: bytes age: 0 date: Mon, 29 Dec 2025 11:01:10 GMT via: 1.1 varnish x-served-by: cache-bom-vanm7210037-BOM x-cache: MISS x-cache-hits: 0 x-timer: S1767006071.720242,VS0,VE223 vary: Accept-Encoding x-fastly-request-id: a9f6fc43381f8bdc776aed4ff5641569e0548b42 content-length: 7986 FAST

SEGA: A Stepwise Evolution Paradigm for Content-Aware Layout Generation with Design Prior

Haoran Wang¹, Bo Zhao², Jinghui Wang¹, Hanzhang Wang³, Huan Yang⁴, Wei Ji², Hao Liu¹ Xinyan Xiao¹

¹Baidu Inc., China ²Nanjing University, Nanjing, China ³Harbin Institute of Technology, China ⁴Kuaishou Technology, China
ICCV 2025 (Highlight)

Paper arXiv Code Data

Abstract

In this paper, we study the content-aware layout generation problem, which aims to automatically generate layouts that are harmonious with a given background image. Existing methods usually deal with this task with a single-step reasoning framework. The lack of a feedback-based self-correction mechanism leads to their failure rates significantly increasing when faced with complex element layout planning. To address this challenge, we introduce SEGA, a novel Stepwise Evolution Paradigm for Content-Aware Layout GenerAtion. Inspired by the systematic mode of human thinking, SEGA employs a hierarchical reasoning framework with a coarse-to-fine strategy: first, a coarse-level module roughly estimates the layout planning results; then, another refining module is leveraged to perform fine-level reasoning regarding the coarse planning results. Furthermore, we incorporate layout design principles as prior knowledge into the module to enhance its layout planning ability. Moreover, we present a new large-scale poster dataset, namely BIG-Poster with rich meta-information annotation. We conduct extensive experiments and obtain remarkable state-of-the-art performance improvement on multiple benchmark datasets.

-->

Approach: SEGA

The training pipeline of SEGA includes two stages:

1) Stage-1: A Coarse-level Estimation (CE) module is trained to predict the layouts from scratch.

2) Stage-2: A Fine-level Refinement (FR) module is initialized by weights of CE module, and generate the final layout based on the results of Stage-1.

Dataset: BigPoster-100K

Our dataset includes a wide range of poster categories in real world scenario such as commercial, event, and product posters.

The hierarchical structure of each poster is preserved.

Category distribution of layout elements.

Comparison to State-of-the-Art Methods

Quantitative Results

Table.1 Experimental results on the PKU and CGL dataset.

Method	PKU						CGL
	Graphic				Content		Graphic				Content
	Ali ↓	Ove ↓	Und_l ↑	Und_s ↑	Read ↓	Occ ↓	Ali ↓	Ove ↓	Und_l ↑	Und_s ↑	Read ↓	Occ ↓
	Non-LLM Based
CGL-GAN (IJCAI, 2022)	-	0.0380	-	0.4800	0.0158	0.1320	-	0.0470	-	0.6500	0.0213	0.1400
LayoutDM (CVPR, 2023)	-	0.1720	-	0.4600	0.0201	0.1520	-	0.0260	-	0.7900	0.0192	0.1270
RALF (CVPR, 2024)	0.0031	0.0095	0.9686	0.8981	0.0138	0.1243	0.0023	0.0059	0.9858	0.9652	0.0180	0.1263
LLM Based
PosterLlama (ECCV, 2024)	0.0036	0.0080	0.9874	0.9497	0.0170	0.1380	0.0022	0.0042	0.9823	0.9463	0.0294	0.2453
SEGA w/o FR 7B	0.0037	0.0052	0.9873	0.9471	0.0150	0.1336	0.0023	0.0032	0.9817	0.9522	0.0298	0.2442
SEGA w/o FR (Ens-2) 7B	0.0035	0.0041	0.9892	0.9673	0.0144	0.1305	0.0021	0.0024	0.9879	0.9657	0.0296	0.2438
SEGA（7B）	0.0035	0.0033	0.9897	0.9731	0.0142	0.1286	0.0020	0.0017	0.9913	0.9782	0.0294	0.2430
GT	0.0036	0.0009	0.9950	0.9444	0.0119	0.1185	0.0023	0.0003	0.9937	0.9402	0.0296	0.2390

Table.2 Experimental results on the Crello dataset.

Method	Rule-based Metrics						Aesthetic Scores
	Graphic			Content			S_DL ↑	S_QL ↑	S_TV ↑	S_IO ↑	S_Mean ↑
	Ali ↓	Ove ↓	Und_l ↑	Und_s ↑	Read ↓	Occ ↓	S_DL ↑	S_QL ↑	S_TV ↑	S_IO ↑	S_Mean ↑
	Non-LLM Based
	FlexDM (CVPR, 2023)	0.0122	0.1139	0.6889	0.5034	0.0516	0.4850	4.563	5.126	4.873	5.239	4.950
LLM Based
PosterLlama (ECCV, 2024)	0.0099	0.0238	0.9204	0.7378	0.0395	0.4041	5.292	5.796	5.263	5.819	5.542
SEGA w/o FR 7B	0.0102	0.0121	0.8206	0.6698	0.0304	0.4002	5.553	6.332	5.693	5.448	5.756
SEGA w/o FR (Ens-2) 7B	0.0100	0.0093	0.8501	0.7202	0.0285	0.3957	5.642	6.418	5.811	5.529	5.850
SEGA 7B	0.0086	0.0040	0.9337	0.8978	0.0282	0.3964	5.792	6.411	5.824	5.708	5.941
SEGA w/o FR 13B	0.0102	0.0093	0.8485	0.7315	0.0271	0.3948	5.923	6.624	6.253	5.991	6.197
SEGA w/o FR (Ens-2) 13B	0.0097	0.0075	0.8715	0.7715	0.0260	0.3874	6.128	6.652	6.058	5.822	6.165
SEGA 13B	0.0095	0.0025	0.9541	0.9270	0.0260	0.3907	6.149	6.745	6.348	6.038	6.320
GT	0.0100	0.0116	0.9643	0.8187	0.0259	0.3797	6.882	7.543	6.863	6.025	-

Qualitative Results

Qualitative Results on PKU dataset

Qualitative Results on CGL dataset

BibTeX

@article{wang2025SEGA,
  author    = {Haoran, Wang and Bo, Zhao and Jinghui, Wang and Hanzhang, Wang and Huan, Yang and Wei, Ji and Hao, Liu and Xinyan, Xiao},
  title     = {SEGA: A Stepwise Evolution Paradigm for Content-Aware Layout Generation with Design Prior},
  journal   = {ICCV},
  year      = {2025},
}

Original Source | Taken Source