404 Lora Helper: A CivitAI Contest Toolkit

🚀 Elevate Your 404 Contest Submissions with Specialized LoRAs

A series of LoRAs designed specifically for the CivitAI 404 Contest.
Each LoRA has been trained with the first ~800 image submissions from the contest, offering diverse ways to enhance your entries.

🌟 Available LoRAs:
1️⃣ LoRA1 - The Standard Bearer

• Trained with CivitAI's trainer on 760 images

• Learning rate: 0.0005, Epoch: 16, Steps: 3200

• Optimizer: AdamW8bit, Base model version: sdxl_base_v1-0

2️⃣ LoRA2 - Lycoris FULL Spectrum

• Featuring a variety of algorithms from the Lycoris FULL suite

• Modules include LohaModule, LoConModule, FullModule, and LokrModule

• Tailored adjustments for both UNet and Text Encoder components

module type table: {'LohaModule': 176, 'LoConModule': 150, 'FullModule': 26, 'LokrModule': 700}
enable_conv = true
# UNet Target Modules and Names
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 Modules and Names
text_encoder_target_module = ["CLIPAttention", "CLIPMLP"]
text_encoder_target_name = []  # "token_embedding" not supported
# Module Algorithm Map
module_algo_map = {
    "CrossAttention": {  # Attention Layer in UNet
        "algo": "lokr",
        "dim": 100000000000,
        "factor": 64
    },
    "FeedForward": {  # MLP Layer in UNet
        "algo": "lokr",
        "dim": 100000000000,  # Trigger full matrix
        "factor": 6
    },
    "ResnetBlock2D": {  # ResBlock in UNet
        "algo": "lora",
        "dim": 64,
        "alpha": 1,
        "use_tucker": true,  # Use tucker decomposition for convolution
        "factor": 8
    },
    "CLIPAttention": {  # Attention Layer in TE
        "algo": "loha",
        "dim": 32,
        "alpha": 1
    },
    "CLIPMLP": {  # MLP Layer in TE
        "algo": "lora",
        "dim": 64,  # Trigger full matrix
        "alpha": 1
    }
}

3️⃣ LoRA3 - The Fusion Quartet

• A dynamic blend of four different LoRAs

• Network dimensions and alpha dynamically resized for nuanced results

• A unique approach for diverse artistic outputs

A Merge of 4 loras trained with Civitai trainer

ss_v2:  "False",
ss_network_dim:  "Dynamic",
ss_training_comment:  "FFusion.AI - Dynamic resize with sv_ratio: 16.0 from 416; ",
ss_network_module:  "networks.lora",
ss_base_model_version:  "sdxl_base_v1-0",
ss_network_alpha:  "Dynamic"

4️⃣ LoRA4 - The Compact 404

• A down-scaled version optimized to 64 dimensions

• Combines the power of multiple LoRAs in a more compact form

• Ideal for streamlined yet rich artistic creations

Another Mash Version downscaled to 64 DIM
 {
ss_network_module:  "networks.lora",
ss_v2:  "False",
ss_base_model_version:  "sdxl_base_v1-0",
ss_network_dim:  "Dynamic",
ss_training_comment:  "FFusion.AI - Dynamic resize with sv_ratio: 64.0 from 369; ",
ss_network_alpha:  "Dynamic"
}

🎨 Fusion Examples:

Experiment with combining different LoRAs for unique effects. For instance:

<lora:FF_404_Inspiration:0.4><lora:404-CIvitAI-lora:1><lora:404FFusionV2:0.71>

📖 Recommended Usage:

Pair these LoRAs with Harrlogos XL
& The 404ra - add-on for Harrlogos!
for enhanced text generation in your 404 project.

Note: These LoRAs are crafted to inspire and assist in the CivitAI 404 Contest. We encourage responsible and creative use to explore the boundaries of AI art.
loras are not supposed to provide out of the box results!!!

For further details and access, visit Civitai 404 Contest page!