Oil Painting Style

Description:

This LORA is trained on ~100 classical oil paintings throughout the past millennium. I selected mainly images that give off the renaissance painting vibe.

The model will primarily affect the art style (obviously making it into an oil painting), but it also will often generate older fashion and somewhat affects the composition to fit that renaissance vibe.

Using the Model:

I recommend using YiffyMix v61 instead of base illustrious/Noob with this model. I do not know why, but it just works really well together when using E621 tags.

Tagging for a oil painting vibe, but without affecting the body/face shape:

Positive:

Oil Painting, {prompt}

Negative:

Chubby

Tags for a complete oil painting vibe:

Positive:

Oil Painting, Realistic, {prompt}

Tags for a woman in her 20s (currently does not work for men):

Positive:

Oil Painting, Human Female, Young Adult, {prompt}

Negative:

Chubby, Child

Training notes:

Version 1: (unreleased) was trained with just "Oil Painting, Realistic" as a tag set on 50 images. This resulted in characters not generating properly and characters either becoming too old or too young. The lora also ended up being too strong: hands got distorted and everything was framed.

Version 2: I tagged the images using 6 tags: "Oil Painting, Realistic, Chubby, Child, Old, Young Adult" to each of the source images and gave a brief natural language description. This normalized the model to not affect the baseline body type, which helped the specified Character Generation significantly. The natural language description seems to have given the Lora a dumping ground for anything it couldn't assign to the above tags, minimizing anatomical distortion and picture frames appearing. In painting could still use some work unfortunately..

Training took 11 epochs using these parameters:

{   
"engine": "kohya",   
"unetLR": 1,   
"clipSkip": 2,   
"loraType": "lora",   
"keepTokens": 0,   
"networkDim": 32,   
"numRepeats": 4,   
"resolution": 1024,   
"lrScheduler": "cosine",   
"minSnrGamma": 5,   
"noiseOffset": 0.1,   
"targetSteps": 3570,   
"enableBucket": true,   
"networkAlpha": 16,   
"optimizerType": "Prodigy",   
"textEncoderLR": 1,   
"maxTrainEpochs": 30,   
"shuffleCaption": false,   
"trainBatchSize": 4,   
"flipAugmentation": true,   
"lrSchedulerNumCycles": 3 
}