PixelPonder
收藏资源简介:
PixelPonder是一个新颖的机制,能够通过从图像级别过渡到补丁级别来细化多视觉条件的组合,从而实现更细粒度的可控生成。
PixelPonder is a novel mechanism that can refine the combination of multiple visual conditions by transitioning from the image level to the patch level, thereby enabling finer-grained controllable generation.
PixelPonder数据集概述
基本信息
- 项目名称: PixelPonder:Dynamic Patch Adaptation for Enhanced Multi-Conditional Text-to-Image Generation
- 项目地址: https://github.com/chfyfr/PixelPonder
- 论文地址: https://arxiv.org/pdf/2503.06684
- 模型发布: 2025/3/11
- 代码发布: 2025/3/11
数据集结构
- 根目录: data/
- 子目录:
- images/
- image_000.png
- image_001.png
- ...
- canny/
- hed/
- depth/
- openpose/
- images/
- 元数据文件: data.jsonl
- 子目录:
元数据格式
json { "image": "path/to/your/data/images/image_000.png", "text": "text1", "canny": "path/to/your/data/canny/image_000.png", "depth": "path/to/your/data/depth/image_000.png", "openpose": "path/to/your/data/openpose/image_000.png", "hed": "path/to/your/data/hed/image_000.png" }
模型下载
- PixelPonder模型: https://huggingface.co/chfyfr/PixelPonder
- FLUX模型: https://huggingface.co/black-forest-labs/FLUX.1-dev
- CLIP模型: https://huggingface.co/openai/clip-vit-large-patch14
许可证
- 模型许可证: FLUX.1 [dev] 和 x-flux 许可证
引用
bibtex @article{pan2025pixelponder, title={PixelPonder: Dynamic Patch Adaptation for Enhanced Multi-Conditional Text-to-Image Generation}, author={Pan, Yanjie and He, Qingdong and Jiang, Zhengkai and Xu, Pengcheng and Wang, Chaoyi and Peng, Jinlong and Wang, Haoxuan and Cao, Yun and Gan, Zhenye and Chi, Mingmin and Peng, Bo and Wang, Yabiao}, journal={arXiv preprint arXiv:2503.06684}, year={2025} }




