consult about the different flux network architecture performance

[sjames40]

i just change the network for flux to latent_in_channels=128
num_layers=26
num_single_layers=0
attention_head_dim=384
num_attention_heads=4
joint_attention_dim=1536
pooled_projection_dim=768
adaln_dim=256
ffn_ratio=4.0
processor_type="default"

axes_dims_rope=(48 168 168)
then the performance is not that good compare to the original model for the id consistency. is that possible to connect with email my email is sjliang941231@gmail.com for further question