Hello! Thank you for the excellent work and for open-sourcing the code.
After reading the paper and the codebase, I have a question regarding the training. I noticed that the selection of the error during training is completely random.
Because of this randomness, it seems highly likely that an error extracted from one domain (e.g., a landscape video) might be injected into a video from a completely different domain (e.g., a video about animals or humans).
Wouldn't this content inconsistency between the source of the error and the target video cause issues or introduce harmful noise during training? I am curious to know if the model is robust to this domain mismatch or if there is an underlying mechanism to handle it.
Looking forward to your insights. Thank you!
Hello! Thank you for the excellent work and for open-sourcing the code.
After reading the paper and the codebase, I have a question regarding the training. I noticed that the selection of the error during training is completely random.
Because of this randomness, it seems highly likely that an error extracted from one domain (e.g., a landscape video) might be injected into a video from a completely different domain (e.g., a video about animals or humans).
Wouldn't this content inconsistency between the source of the error and the target video cause issues or introduce harmful noise during training? I am curious to know if the model is robust to this domain mismatch or if there is an underlying mechanism to handle it.
Looking forward to your insights. Thank you!