🌟 Trained specifically for style mixing on Safe For Work (SFW) dataset.
🔹 Text Encoder (strongly reliant on U-Net architecture):
🔹 Lora Model:
Trained using the CivitAI trainer.
Underwent training for 16 epochs.
Performance: Lora showcases stable diffusion in XL settings.
🔹 Lycoris Models:
Configured using this preset from LyCORIS, but with our custom tweaks.
Trained on 8xA6000 GPUs.
Underwent 20-40 epochs with a batch size of 20.
Module Type Breakdown:
LohaModule: 176
LoConModule: 150
FullModule: 26
LokrModule: 700
🏆 Comparison:
CivitAI's Lora stands out in our tests, but we've yet to explore its full potential in style mixing with other models.
Feedback, collaboration, and tests are welcome! 🤟 🥃
Description
Mixed training experiment using | lokr lora loha
More balanced dims (lower size)
module type table: {'LohaModule': 176, 'LoConModule': 150, 'FullModule': 26, 'LokrModule': 700}
@unet_target_module = [
"Transformer2DModel",
"ResnetBlock2D",
"Downsample2D",
"Upsample2D",
]
unet_target_name = [
"conv_in",
"conv_out",
"time_embedding.linear_1",
"time_embedding.linear_2",
]
text_encoder_target_module = [
"CLIPAttention",
"CLIPMLP",
]
[module_algo_map]
[module_algo_map.CrossAttention] # Attention Layer in UNet
algo = "lokr"
[module_algo_map.FeedForward] # MLP Layer in UNet
algo = "lokr"
[module_algo_map.ResnetBlock2D] # ResBlock in UNet
algo = "lora"
[module_algo_map.CLIPAttention] # Attention Layer in TE
algo = "loha"
[module_algo_map.CLIPMLP] # MLP Layer in TE
algo = "lora"
FAQ
Comments (1)
Some of the images really need some work.