SEGA

<center><font size=4>[ICCV 2025 Spotlight] SEGA: A Stepwise Evolution Paradigm for Content-Aware Layout Generation with Design Prior</font></center>

<img src="Pull_Figure.png" width="750" height="350">

Updates

• [02/11/2026] We release the GenPoster-100K dataset on Hugging Face Website.
• [11/24/2025] We release the inference code and models.
• [10/25/2025] We release the GenPoster-100K dataset.
• [06/25/2025] SEGA is accepted by ICCV 2025.

Code

Envs

• cuda_12.x
• python=3.10
• pytorch 1.7.1
• pip install -r requirements.txt
• python -m spacy download zh_core_web_sm

Dataset

We provide the PKU-style Crello dataset for quick training and inference.

Link: https://pan.baidu.com/s/1hAyl_LatXIW-PEXnXzRCcw password: 1234

Models

We upload all Crello related Models by Baidu Netdisk. (You only need part of them. Below is specific path.)

Link: https://pan.baidu.com/s/1jW7jMjWEOWCgSTU-jUjsNw password: 1234

Below, if no specific notation, all are Lora folder.

7B Model

• SFT model : zzz_git/sft
• Refiner from Llava : zzz_git/refine

<!-- - SFT model merged : -->

13B Model

• SFT merged: base_sft.tar
• SFT merged pretrained: big.tar
• SFT model : crello_series1/simple_sft_17e_nocot_all_old
• Refiner from Bigposter : crello_series2/refine_final_11_1_10e_all_from_bigposter (--base use SFT merged pretrained)
• Refiner from SFT : crello_series2/refine_final_11_1_10e_from_sft (--base use SFT merged)
• Refiner from Llava : crello_series2/ refine_final_11_1_10e_basedata_nocot400_70p

Inference

Preparetion

Download openai/clip-vit-large-patch14-336 in SEGA dir.

Download llava-1.5 7B and 13B for use load checkpoints.

Download Awesomeposter for utils and fonts by link and place it in SEGA dir : https://pan.baidu.com/s/1XoTwYEPbW3rsoKN-VtxOtw 提取码: 1234

sfonts.init('./AwesomePoster/fonts') ( Focus this path in inference script)

<!-- sys.path.append("/data1/zb/LLaMA-Factory-main") -->

Use

• py_folder/zzz_infer_crello.py : SFT inferenc script
• py_folder/sft_infer_crello.py : Refine inferenc script

Here is an example:

python zzz_infer_crello.py --out "/data1/zb/LLaMA-Factory-main/temp_out/debug" --card "6" --lora "simple_sft_17e_nocot_all_old" --base  /data1/zb/ckpts/llava13b &
python zzz_infer_crello.py --out "/data1/zb/LLaMA-Factory-main/temp_out/debug" --card "7" --lora "simple_sft_17e_nocot_all_old" --base /data1/zb/ckpts/llava13b 

BibTeX

If you find this work helpful, please cite our work:

@inproceedings{wang2025sega,
  title={SEGA: A Stepwise Evolution Paradigm for Content-Aware Layout Generation with Design Prior},
  author={Wang, Haoran and Zhao, Bo and Wang, Jinghui and Wang, Hanzhang and Yang, Huan and Ji, Wei and Liu, Hao and Xiao, Xinyan},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={19321--19330},
  year={2025}
}