Royalvice / DocDiff

ACM Multimedia 2023: DocDiff: Document Enhancement via Residual Diffusion Models. Also contains 1597 red seals in Chinese scenes, along with their corresponding binary masks.
https://www.aibupt.com/
MIT License
196 stars 21 forks source link

推理时长? #6

Closed samaritan1998 closed 10 months ago

samaritan1998 commented 10 months ago

请问一下推理时长有测过么,和有和Gan对比过嘛?

Royalvice commented 10 months ago

可参考原文表5,每推理100万像素图片(包括IO),默认配置下DocDiff(Non-native)-5耗时0.33秒,DocDiff(Non-native)-100耗时5.69秒,MPRNet耗时0.57秒,DE-GAN耗时0.82秒。总体来说,DocDiff-5速度还是很可观的,因为DocDiff的HRR模块只有4M,即使采样5次但是计算时间复杂度也很低。但是DocDiff-100速度还是很慢的。不过,DocDiff在文档图像上的训练十分稳定。如果跑过DE-GAN的官方代码,可以知道GAN方法很难收敛。

Royalvice commented 10 months ago

如果您还有什么问题,欢迎联系我。如果您觉得有帮助并给个Star,我竟感激不尽,谢谢!

Royalvice commented 10 months ago

请问一下推理时长有测过么,和有和Gan对比过嘛?

Demo下有notebook,您可以自己测一测速度

Royalvice commented 10 months ago

可参考原文表5,每推理100万像素图片(包括IO),默认配置下DocDiff(Non-native)-5耗时0.33秒,DocDiff(Non-native)-100耗时5.69秒,MPRNet耗时0.57秒,DE-GAN耗时0.82秒。总体来说,DocDiff-5速度还是很可观的,因为DocDiff的HRR模块只有4M,即使采样5次但是计算时间复杂度也很低。但是DocDiff-100速度还是很慢的。不过,DocDiff在文档图像上的训练十分稳定。如果跑过DE-GAN的官方代码,可以知道GAN方法很难收敛。

单张3090