This one was more than tricky to make! The key is using a low tagging, as you don't want the style to be tied to specific keywords. Not tagging at all (excepted trigger words) should have worked too.
Trained with OneTrainer on public domain paintings provided by the National Gallery of Art.

• 173 paintings tagged "Baroque"
• 133 paintings tagged "Impressionism"
• 87 paintings tagged "Realism" (Repeated 2x during the training)
• 70 paintings tagged "Rococo" (Repeated 2x during the training)
• 144 paintings tagged "Romanticism"

Steps

1. Download images
2. Resize them: I made a PHP script to resize images with a minimum size of 300 pixels and a maximum of 576, converts them to PNG format and adding a prefix for better classification.
   See the Code - Download it
   Dimensions are grouped by multiples of 8 to facilitate batch processing.
   Will give a try with cropping too, as resizing can lead to severally low resolution.
   <?php
       
$dest_rep = 'to_crop/output/'; // Output directory
$basefiles = 'to_crop/input/'; // Input directory
$files = scandir($basefiles);
$count = 0;
$max_size = 576; // Max dimension

foreach($files AS $useless => $name)
{
    if($name != '.' && $name != '..')
    {
        $source_file = $basefiles . $name;
        $dest_file = $dest_rep . 'painting_romanticism_' . $count . '.png'; // Change 'painting_romanticism_' for the prefix you want
        
        if(substr($name,-4) == '.png')
        {  
            $img = @imagecreatefrompng($source_file);
        }
        elseif(substr($name,-4) == '.jpg' || substr($name,-5) == '.jpeg')
        {  
            $img = imagecreatefromjpeg($source_file);
        }
        elseif(substr($name,-5) == '.webp')
        {  
            $img = imagecreatefromwebp($source_file);
        }
        elseif(substr($name,-5) == '.avif')
        {  
            $img = imagecreatefromavif($source_file);
        }           
        
        $original_x = imagesx($img);
        $original_y = imagesy($img);
       
        // Optimal ratio
        $ratio = ($original_x / $original_y) > 1 ? 169 : 916; 
        $new_x = floor($original_x / 8) * 8;
        $new_y = floor($original_y / 8) * 8;
        
        if($ratio == 169)
        {
            $true_ratio = $original_x / $original_y;                    
            if($original_x > $max_size)
            {
                $ratio_crop = $original_x / $max_size;
                $new_x = $max_size;
                $new_y = round($original_y / $ratio_crop);
                $new_y = floor($new_y / 8) * 8;
            }                    
        }
        else
        {
            $true_ratio = $original_y / $original_x;
            if($original_y > $max_size)
            {
                $ratio_crop = $original_y / $max_size;
                $new_y = $max_size;
                $new_x = round($original_x / $ratio_crop);
                $new_x = floor($new_x / 8) * 8;
            }
        } 

        if($new_x > 300 && $new_y > 300)   
        {    
            $dst_file = imagecreatetruecolor($new_x, $new_y);
            if(substr($name,-4) == '.png' || substr($name,-5) == '.webp')
            {  
                imagecolortransparent($dst_file, imagecolorallocatealpha($dst_file, 0, 0, 0, 127));
                imagealphablending($dst_file, false);
                imagesavealpha($dst_file, true);
            }               
            imagecopyresampled($dst_file, $img, 0, 0, 0, 0, $new_x, $new_y, $original_x, $original_y); 
            imagepng($dst_file, $dest_file);  
        }    
    }
}

?>    
3. Automatic tagging with Kohya_SS, using WD14 and a treshold of 0.8, adding "paintify" + painting style at the beginning
   Treshold is high because I want the style to be applied globally.
4. I Use OneTrainer for the training: download settings.
   As this is a style LoRA, learning rate is slowed and steps are increased.