How to train EDLoRA with higher-resolution images?

[volcverse]

Hello! Thanks for your great work! Recently I'm trying to train edlora with high-resolution images, 1024 x 1024. I noticed that the results with new_concept_embedding are not good. They are very grainy. While the results without new_concept_embedding are good but can't preserve the subject identity because of lack of new_concept_embedding. So I wonder what I should do to train a edlora with higher resolution images?

Currently, my training settings are:
change latent_size to [4, 128, 128]
instance_transform:
- { type: Resize, size: 1024}
comment the line below:
# - { type: HumanResizeCropFinal, size: 1024, crop_p: 0.5 }
All other settings remain default as the example config. And the sample image resolution in the code has already been changed to 1024x1024

[guyuchao]

Could you mind share your concept data with me ([email protected]), I will debug it and update the code accordingly. I haven't tried on 1024^2 image yet.

[volcverse]

Could you mind share your concept data with me ([email protected]), I will debug it and update the code accordingly. I haven't tried on 1024^2 image yet.

Hello! Greatly appreciate your response and an email is sent!