CivitAI Style Fusion🏆LoRAs

Last update: 🚀 CivitAI Lora5 32DIM Notebook with dataset

Last update: 🚀 CivitAI Lora3 Configuration - Trained with CivitAI Trainer

🚀 Date: 2023-11-10 | Title: CivitAI_??_ALL

🔍 Key Specifications:

Resolution: 1024x1024
Architecture: stable-diffusion-xl-v1-base/lora
Network Dim/Rank: ??.0, Alpha: 1.0
Module: networks.lora
Learning Rates: UNet LR & TE LR set to optimal levels
Optimizer: Advanced AdamW8bit
Epochs & Training: Intensive 10 epochs with 576 batches

📊 Model Stats:

UNet Weight: Mag - 7.602, Str - 0.0187

Resolution: 1024x1024 Architecture: stable-diffusion-xl-v1-base/lora
Network Dim/Rank: ??.0 Alpha: 1.0
Module: networks.lora
Learning Rate (LR): 0.0005 UNet LR: 0.0005 TE LR: 5e-05
Optimizer: bitsandbytes.optim.adamw.AdamW8bit(weight_decay=0.1)
Scheduler: constant  Warmup steps: 0
Epoch: 10 Batches per epoch: 576 Gradient accumulation steps: 1
Train images: 2304 Regularization images: 0
Multires noise iterations: 6.0 Multires noise discount: 0.3
Min SNR gamma: 5.0 Zero terminal SNR: True Max grad norm: 1.0  Clip skip: 1
Dataset dirs: 1
        [img] 576 images
UNet weight average magnitude: 7.602270778898858
UNet weight average strength: 0.018722912685324843
Text Encoder (1) weight average magnitude: 2.7??9271326702607
Text Encoder (1) weight average strength: 0.009535635958680934
Text Encoder (2) weight average magnitude: 2.6905091182810352
Text Encoder (2) weight average strength: 0.007233532415344915

Delve into FFusionAI's approach to AI-driven style synthesis with our newly released LoRA models. Each model has been developed using CivitAI's official trainer, ensuring precision and quality.

🛠️ LoRA Model Overview:

LoRA 1 - Lite Version: Designed for quick testing, this model utilizes a small dataset for swift style generation, operating with a 32-dimension capacity.
LoRA 2 - Community Fusion: A robust model developed from over 500+ images, submitted by various users for the CivitAI contest. This iteration also features a 32-dimension capacity.
LoRA 3 - Enhanced Fidelity: Building upon LoRA 2, this model is further trained with higher dimensions, focusing on improving the overall image quality.
LoRA 4 - Comprehensive Style Mash: Our expansive dataset of 1400 images represents a confluence of all FFusionAI submissions. This model undergoes additional UNet training to refine and diversify the generated styles.

1. FFusionAI Style Capture & Fusion Showdown LoRA

🎨 Dataset and Training:

Included within the package are curated collections accessible at CivitAI Collections. The training prompts have been crafted with BLIP-2, FLAN-T5-XL, and ViT-H-14.

Please note, original prompts were not utilized for training. Instead, intentional modifications were made using blip2-flan-t5-xl & ViT-H-14/laion2b_s32b_b79k to adjust and enhance the training dataset, which can be reviewed here.

🔄 Further Information:

For a detailed examination of the training datasets, parameters, and model specifications, professionals and enthusiasts are encouraged to explore the metadata provided within the collection.