I feel like it’s similar to image compression, you lose a bit every iteration. Consider that the original model was weighted towards common aspects across the training set. Even with some creative prompting for your source images you could unintentionally introduce bias and reduce variations across images generated by your new model. You also get any mistakes or inconsistencies baked in.
The whole “don’t train on synthetic data” thing only holds true if you want to represent real world data, and even then it can be extremely useful.
But if the synthetic data is what you’re trying to replicate anyway, train away!
I feel like it’s similar to image compression, you lose a bit every iteration. Consider that the original model was weighted towards common aspects across the training set. Even with some creative prompting for your source images you could unintentionally introduce bias and reduce variations across images generated by your new model. You also get any mistakes or inconsistencies baked in.
As long as the distortions aren’t noticeable, no one can complain.
I didn’t know that. People make it seem like your model will explode if you do that.
Yep. What’s the worst that can happen? You bias your model towards what you’re actually trying to achieve?
It’s really only a problem for targeting realism, because generated images simply aren’t real.