📌 influence_girl

Learning Rates:

• unetLR: 0.00035

• textEncoderLR: 0.00015

Clip Skip: 2 (helps avoid over-saturation or repetition in certain text encodings)

LoRA Settings:

• loraType: “lora”

• networkDim: 32

• networkAlpha: 16 (maintains quality with reasonable VRAM usage)

Training Strategy:

• numRepeats: 4

• keepTokens: 1

• shuffleCaption: true (improves generalization)

• flipAugmentation: false (avoids excessive mirroring)

Resolution: (Specify your desired pixel resolution, e.g., 512×512 or 768×768)

Scheduler: cosine_with_restarts, with lrSchedulerNumCycles: 3

Noise & Regularization:

• minSnrGamma: 5

• noiseOffset: 0.15

Target Steps: 5400

Max Epochs: 12

Training Batch Size: 2