Training Details of Cyclegan_turbo from Night to Day？

[weihongwei-zg]

Hello author, I noticed that the night to day effect in this project is very stunning. I would like to reproduce this work and use it in my own data scenario. But the document only includes the conversion of zebras and does not include the details of day and night conversion. Can you provide the amount of data for day and night conversion, as well as the training configuration?

[GaParmar]

We use the BDD100k driving dataset for the day to night conversion model.
For dataset processing, we resize all images to 512x512 during training and inference.
The other configurations should be the same!

[weihongwei-zg]

Hello, my training set consists of 6000 paired images that generate visible light from infrared. The training environment is a 4090 graphics card, The batch_size is set to 1, with a size of 256x256. After 3 iterations, the effect is poor and not as good as the original cyclegan algorithm. Could you please help to point out the reason for this?

[weihongwei-zg]

It seems that the brightness and darkness are opposite.

[kaka45inablink]

@GaParmar Hi dear author,
May I ask how to set the prompt for the day to night conversion model when training?
"driving in the night" for a2b and "driving in the day" for b2a?