hans1984 / glowgan Goto Github PK

Hello!

Thank you for your original and inspiring work. I have a small question regarding the technical details. During adversarial training, is it necessary for the real LDR and the generator-produced LDR to have consistent or similar exposure levels when calculating the loss? In your paper, you mentioned that you did not retrieve the EXIF data of the LDRs. If the generated LDRs have randomly sampled exposure values, in extreme cases, the discriminator might receive one LDR with a very long exposure time and another with a very short exposure time. Can the training still ensure stability under such circumstances?

Thanks for your great work! Could you provide the LDR dataset for training? It will be pretty helpful.