search

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

Inference time #3

Closed Ru-Van closed 1 year ago

Ru-Van commented 1 year ago

Can you please add inference time information for different configurations and video cards?

Royalvice commented 1 year ago

For the network under the default conf.yml configuration, inferring one million pixels on a single RTX3090 takes about 5 seconds, sampling 100 times. I haven't conducted more comprehensive experiments on inference times. FLOPs and MACC haven't been calculated, but since the network structure of DocDiff is a general and tiny Unet, I believe there are many methods for calculation and testing available on the Internet.