Open pureexe opened 2 months ago
With the Inversion, we found that the condition is somehow reconstructed back to the dog, even though the network didn't feed the dog. Probably this is overfitting on training.
After the discussion we may try to use depth or normal to help network to disentangle the lighting data.
I confirmed that this is not the input bug. The model already see the corrected input during DDIM inversion
The previous comment is incorrect. It actually the input bug. but it bug on the prompting part
Update: with fixed prompt: the image still doesn't relight (Source image show light coming from left direction while target light is coming from the right)
Update: Our code is now already support the Inversion
Since we discussed on #4
We can use inversion code to inverse network