🗡️ FFusionAI's Knife LoRA Model Demonstrations

This demonstration dives deep into the intricacies of three distinct LoRA trainings. Each model has been meticulously trained on a unique dataset of 200 knives, captured in the professional environment of our partner, NoramePhotography Studio. The visuals span from pure white studio shots to rustic wood settings, glinting coins, and fantasy-inspired indoor decor.

🔍 Dataset Insight: The dataset, although rich in variety, was curated with a fast and informal tagging approach, mainly for demonstration purposes. If you're intrigued by the knife photo session and wish for a more in-depth training, do let us know!

While depth variations are in the pipeline, our current focus revolves around evaluating the distinct LoRA variations.

🎯 Models at a Glance:

1. CivitAI's Quick LoRA Training (Lora1)

📌 Highlights:

Powered by CivitAI's new LoRA trainer.
Swift 10-epoch run, completed in a breezy 20-30 minutes.
Quality may vary with default settings, but hey, time is essence!

📊 Specifications:

Date: 2023-09-19T14:36:14
Resolution: 1024x1024
Architecture: stable-diffusion-xl-v1-base/lora
Network Dim/Rank: 32.0
Alpha: 16.0

Knife_XL_FFusion.safetensors
Date: 2023-09-19T14:36:14 Title: Knife_XL_FFusion
Resolution: 1024x1024 Architecture: stable-diffusion-xl-v1-base/lora
Network Dim/Rank: 32.0 Alpha: 16.0
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: cosine_with_restarts  Warmup steps: 0
Epoch: 10 Batches per epoch: 74 Gradient accumulation steps: 1
Train images: 282 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] 282 images
UNet weight average magnitude: 2.634092236933176
UNet weight average strength: 0.009947009810559605
Text Encoder (1) weight average magnitude: 1.696394163771355
Text Encoder (1) weight average strength: 0.008538951936953606
Text Encoder (2) weight average magnitude: 1.720911101275907
Text Encoder (2) weight average strength: 0.006699097931942388

2. LoRA FA with Text Encoder Only (Lora2)

📌 Highlights:

Exclusive training on text encoder.
Absence of UNet in this LoRA variant.

📊 Specifications:

Date: 2023-09-19T20:04:24
Resolution: 1024x1024
Architecture: stable-diffusion-xl-v1-base/lora
Network Dim/Rank: 32.0
Alpha: 32.0

Knife-FFusion-LoRA-FA.safetensors

Date: 2023-09-19T20:04:24 Title: Knife-FFusion-LoRA-FA
Resolution: 1024x1024 Architecture: stable-diffusion-xl-v1-base/lora
Network Dim/Rank: 32.0 Alpha: 32.0
Module: networks.lora_fa

Text Encoder (1) weight average magnitude: 3.986337637923385
Text Encoder (1) weight average strength: 0.018590648076750333
Text Encoder (2) weight average magnitude: 4.043434837883338
Text Encoder (2) weight average strength: 0.014620680042179104
No UNet found in this LoRA

3. General LoRA Training

📌 Highlights:

Comprehensive LoRA training with diverse specifications.
Trained on an extensive dataset of 485 knife images.

📊 Specifications:

Date: 2023-08-26T23:08:56
Resolution: 1024x1024
Architecture: stable-diffusion-xl-v1-base/lora
Network Dim/Rank: 32.0

Alpha: 16.0

FF-Minecraft-XL
Resolution: 1024x1024 Architecture: stable-diffusion-xl-v1-base/lora
Network Dim/Rank: 32.0 Alpha: 16.0
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: cosine_with_restarts  Warmup steps: 0
Epoch: 10 Batches per epoch: 121 Gradient accumulation steps: 1
Train images: 458 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] 458 images
UNet weight average magnitude: 2.9987627096874507
UNet weight average strength: 0.011098071585284945
Text Encoder (1) weight average magnitude: 1.729993708156961
Text Encoder (1) weight average strength: 0.008685239007756952
Text Encoder (2) weight average magnitude: 1.7630326984758309
Text Encoder (2) weight average strength: 0.0068346636309082635

🎨 Readme Crafted by: 🤖 & FFusionAI 🚀

🌐 Contact Information

The FFusion.ai project is proudly maintained by Source Code Bulgaria Ltd & Black Swan Technologies.

📧 Reach us at di@ffusion.ai for any inquiries or support.

🌌 Find us on:

🌍 Sofia Istanbul London

Knife XL FFusion - CivitaI / LoRA + FA Text Encoder - Knife XL FFusion CivitAI

Knife XL FFusion - CivitaI / LoRA + FA Text Encoder - Knife XL FFusion CivitAI

Sampler

Steps

CFG Scale

Seed

Clip Skip

Size