This is a brief disclosure and look at what went into the "Dark Haunted Fantasy" LORA
Method:
To construct the dataset, high-resolution images were meticulously edited and combined to represent specific elements of the concept to be conveyed.
Scaling:
To ensure the absence of undesirable elements, meticulous care was exercised during the scaling process. Manual scaling was employed.
When scaling an image to less than half of its original size to achieve the final 1024*1024 size, a sigmoidized EWA (Elliptical Weighted Averaging) with the Robidoux bicubic sampling technique was blended with a non-sigmoidized EWA Robidoux. Otherwise, a nonlinear interpolation was used to create a double-density version of the original image; this double-density image was then resampled with bilinear interpolation. This process was undertaken to eliminate the effects of rounding error and achieve monotonicity, thereby maximizing the dynamic range of the image.
Training Parameters:
Given the nature of the images (we are not training the Lora to draw specific objects or people in the images), and many images share no common objects with any other image, a batch size of 1 was utilized, and repetition was not employed. Prodigy was employed as the optimizer. The learning rate was set to a constant value of 1, and the network alpha was set to be equal to the network dimension, which was 16 in this case.
The LoRA model was trained using the Civitai on-site LORA Trainer.
The following are the raw settings utilized by the Civitai on-site LORA Trainer:
{
"engine": "kohya",
"unetLR": 1,
"clipSkip": 1,
"loraType": "lora",
"keepTokens": 0,
"networkDim": 16,
"numRepeats": 1,
"resolution": 1024,
"l": "constant",
"minSnrGamma": 5,
"noiseOffset": 0.1,
"targetSteps": 7560,
"enableBucket": true,
"networkAlpha": 16,
"optimizerType": "Prodigy",
"textEncoderLR": 0,
"maxTrainEpochs": 90,
"shuffleCaption": false,
"trainBatchSize": 1,
"flipAugmentation": false,
"lrSchedulerNumCycles": 1
}
Dataset Size:
The dataset in version 5 contains 84 Images and labels all scaled to 1024*1024.
Sample images from the Dataset:
