Training Params
Expand
{
"engine": "kohya",
"unetLR": 0.0005,
"clipSkip": 2,
"loraType": "lora",
"keepTokens": 3,
"networkDim": 16,
"numRepeats": 20,
"resolution": 1024,
"lrScheduler": "cosine_with_restarts",
"minSnrGamma": 5,
"noiseOffset": 0.1,
"targetSteps": 9945,
"enableBucket": true,
"networkAlpha": 14,
"optimizerType": "Adafactor",
"textEncoderLR": 0.00005,
"maxTrainEpochs": 13,
"shuffleCaption": true,
"trainBatchSize": 4,
"flipAugmentation": true,
"lrSchedulerNumCycles": 3
}
Description
[
{"prompt": "married malay men in a group showing off"},
{"prompt": "grandfather"},
{"prompt": "thick hairy pubes,"}
]
training a custom AI image generation model using a technique called LoRA (Low-Rank Adaptation). Here's what each section means in plain terms:
Network Architecture Settings
What it does: Defines how the AI learns new visual concepts without completely retraining the entire model.
Learning rates (0.0005 for image generation, 0.00005 for text understanding): How quickly the AI adapts - like adjusting how big steps you take when learning to ride a bike
Network dimensions (16) and alpha (14): Control how much the model can change - smaller numbers mean more subtle adjustments
LoRA method: A memory-efficient way to teach new concepts to existing AI models
Usage example: Training a model to generate images of a specific character, art style, or object while keeping the base model's general knowledge intact.
Learning Optimization
What it does: Controls how the AI improves during training, like setting study schedules and methods.
Cosine scheduler with restarts: The learning intensity follows a wave pattern, starting strong, tapering off, then restarting - like interval training
Adafactor optimizer: An efficient method for updating the model that uses less memory
13 training cycles: The AI will see your training images 13 times to learn the patterns
Usage example: Training a model to recognize your pet's specific features across different poses and lighting conditions.
Training Process Controls
What it does: Manages the practical aspects of training - batch sizes, memory usage, and quality settings.
Batch size 4: Processes 4 images simultaneously for efficiency
Noise offset 0.1: Adds slight randomness to prevent the model from memorizing exact images
Mixed precision: Uses less memory while maintaining quality
Gradient checkpointing: Trades computation time for memory savings
Usage example: Fine-tuning settings for your hardware - smaller batches for limited GPU memory, larger batches for powerful systems.
Advanced Quality Features
What it does: Implements sophisticated techniques to improve training stability and output quality.
Multi-resolution noise: Helps the model learn details at different scales
Min SNR gamma: Prevents the model from focusing too much on very noisy training examples
Clip skip: Adjusts how the text and image understanding components interact
Usage example: Creating a model that can generate both close-up details and wide landscape shots of the same subject with consistent quality.
Practical Applications
This configuration would be ideal for:
Character consistency: Training a model to generate the same fictional character across different scenes
Art style replication: Teaching an AI to mimic a specific artist's technique
Product visualization: Creating variations of a product for marketing materials
Concept art development: Generating multiple iterations of a design concept
The settings are balanced for moderate hardware requirements while maintaining good quality output, making it suitable for hobbyists and small studios rather than requiring enterprise-level computing resources.