LewdXL - V-Pred Finetune

Artists:

• rolua

• deadpurity (AI Artist)

• Pluvium_grandis

• murai-renji / range murata

• baifeidaiwang

• chen_bin

• hanozuku

• 8rk9

• narue

• akifn

• hatena (nazequestion)

• dor_m

• gagaimo

• sweetonedollar

• popepopo999

⚙️ Training Configuration

Core Parameters

• Base Model: Zeronansv9 (Noobai based)/Stable Diffusion XL (SDXL)

• Prediction Mode: V-Prediction (force_v_prediction: true)

• Training Epochs: 50

• Training Steps: 10150

• Batch Size: 8

• Learning Rate: 6e-6

• Text Encoder LR: 2e-6

• Resolution: 1024x1024

Optimization Settings

• Scheduler: 1-cycle cosine

• Warmup Steps: 100

• Min LR Factor: 0.1

• Gradient Accumulation: 1

• Gradient Norm Clipping: 1.0

• Dropout: 0.05

• Precision: BF16

Loss Configuration

• MSE Loss: 1.0

• MAE Loss: 0.1

• CosH/VB Loss: 0.0

• Constant Loss Weight: 2.0

🧠 Infrastructure & Training Setup

• Trainer Framework: OneTrainer

• Platform: RunPod (Community Cloud)

• Hardware: L40S GPU

• Device: CUDA

• EMA: Enabled (decay: 0.999, update_interval: 5)

• Sync Method: SCP (native file sync)

• Autocast Caching: Enabled

🔧 Model Components

All major components were actively trained:

• UNet: ✅ Trained

• Text Encoders (4x): ✅ Trained

• VAE: ✅ Trained

• Prior: ✅ Trained

• EffNet/Decoder/VQGAN: ✅ Trained

Additional Settings

• Masked Training: Disabled

• Offset Noise Weight: 0.05

• Perturbation Noise Weight: 0.05

• Timestep Distribution: Uniform [0.0 → 1.0]

📦 Release Notes

This is the initial alpha release (v0.1) of LewdXL, focused on establishing baseline anime style reproduction with V-Prediction architecture. This experimental checkpoint tests core functionality and artistic style transfer capabilities.

Expected Improvements in Future Versions:

• Enhanced prompt adherence and control

• Improved image coherency and consistency

• Greater output diversity and variation

• Refined artist style blending

• More artists, concepts and characters