🤖 AI Model Training Summary

This document details the configuration and results of your completed Low-Rank Adaptation (LoRA) training job, which customized the Stable Diffusion XL (SDXL) model.

1. Job Status and Timeline

• Training Start: $\text{October 18, 2025 at 06:39:35 PM}$

• Completion Time (Ready): $\text{October 19, 2025 at 04:17:36 AM}$

• Total Duration: Approximately 9 hours and 38 minutes.

• Job ID: $\text{563759-20251018103932646}$

• Base Model Used: SDXL (Stable Diffusion XL)

History Milestones:

• Submitted: $\text{10/18/2025 06:39:33 PM}$

• Processing: $\text{10/18/2025 06:39:35 PM}$

• Ready (Complete): $\text{10/19/2025 04:17:36 AM}$

2. Dataset and Configuration

• Dataset Size: $\text{27}$ image files.

• Label Count: $\text{27}$ corresponding descriptions (one for each file).

• Label Type: Tags (using short keywords for description).

• Training Engine: Kohya (a popular training utility).

• LoRA Type: lora (standard Low-Rank Adaptation).

• Privacy: $\text{Own Rights}$ (You hold the rights to the training data).

3. Key Training Parameters (The Recipe)

These settings determined how the model learned from your $\text{27}$ images.

Learning Control

• Maximum Epochs: $\text{20}$ (The dataset was shown to the model 20 times).

• Target Steps: $\text{9,990}$ (The total number of learning batches aimed for).

• Num Repeats: $\text{74}$ (This increases the total steps to ensure the model focuses heavily on your small dataset).

• Train Batch Size: $\text{4}$ (The model processed 4 images at a time during each step).

• Shuffle Caption: $\text{true}$ (The order of tags was randomized to improve learning).

Learning Rates (Speed of Learning)

• U-Net Learning Rate (unetLR): $\text{0.0005}$ (Rate for the image generation part).

• Text Encoder Learning Rate (textEncoderLR): $\text{0.00005}$ (Rate for the text understanding part—slower is typical).

• Optimizer Type: Adafactor.

• LR Scheduler: cosine_with_restarts (Determines how the learning rate changes over time).

Model Size and Image Processing

• Resolution: $\text{1024}$ (Training size in pixels, standard for SDXL).

• Network Dimension (networkDim): $\text{32}$ (A measure of the LoRA's capacity or strength).

• Network Alpha (networkAlpha): $\text{16}$ (Used with networkDim for LoRA stability).

• Enable Bucket: $\text{true}$ (Helps handle images of slightly different aspect ratios efficiently).