📂 Masha K v3 ‑ root — Parameter Settings

🔧 Network Module

• Type: LoRA

• Base Model: FLUX – Dev fp32

• Trigger words: maxxa

🖼️ Image Processing Parameters

• Repeat: 8

• Epochs: 8

• Save Every N Epochs: 1

⚙️ Training Parameters

• Seed: 42

• Clip Skip: 1

• Text Encoder LR: 0.00002

• UNet/DiT LR: 0.0002

• LR Scheduler: constant

• Optimizer: AdamW

• Network Dim: 100

• Network Alpha: 128

• Gradient Accumulation Steps: 4

🏷️ Label Parameters

• Shuffle caption: –

• Keep n tokens: –

🧪 Advanced Parameters

• Noise offset: 0.008

• Multires noise discount: 0.06

• Multires noise iterations: 4

• conv_dim: 128

• conv_alpha: 160

• Batch Size: –

📸 Sample Image Settings

• Prompt:

  maxxa. A blonde woman stands confidently in a forest with ethereal atmosphere. Her body softly accentuated by directional lighting. Her long wavy hair cascades past her shoulders with strand realism. Her blue eyes are enhanced with mascara and subtle eyeliner, paired with red lips, a small nose, and a calm, focused expression. She is wearing a pink top, red leather skirt and red stockings. Her posture is upright and poised, one hand resting on her waist, the other hand on her thigh. Her legs are crossed. She is looking at the viewer, daring expression.

• Sampler: dpmpp_2m

✨ This layout keeps each block separated and easy to scan — perfect for logging or archiving.

Do you want me to also create a repeat distribution table (x1, x2, x3 counts for wides, fulls, close‑ups) in the same style so you can track dataset balance alongside these parameters?