The Code I Use To Maximize the Effectiveness of my Pony LoRAs.

So, you're probably wondering what sort of data prep I do for my Pony LoRAs. Well, it's the same as I do for my other LoRAs (basically the Kohya method) but with these two codes:

A code to delete images that don't fit inside the smallest ratio that fits SDXL.

from PIL import Image

# Set the paths to your photo folder
photo_folder = "/content/drive/MyDrive/Loras/[INSERT FOLDER HERE]/dataset"

# Define the minimum size threshold in pixels (width * height)
min_size_threshold = 983040  # Adjust this value as needed

# Initialize counters
photos_deleted = 0
photos_remaining = 0

# Iterate through each photo file in the folder
for photo_file in os.listdir(photo_folder):
    # Check if the file is an image
    if not photo_file.lower().endswith(('.jpg', '.jpeg', '.png')):
        continue  # Skip non-image files

    # Get the photo size using Pillow
    photo_path = os.path.join(photo_folder, photo_file)
    try:
        with Image.open(photo_path) as img:
            width, height = img.size
            photo_size = width * height
    except OSError:
        # Handle potential errors while opening images
        print(f"Error opening {photo_file}. Skipping...")
        continue

    # Check if the size is below the threshold
    if photo_size < min_size_threshold:
        # Delete the photo
        os.remove(photo_path)
        print(f"Deleted: {photo_file} (Size: {photo_size})")
        photos_deleted += 1
    else:
        photos_remaining += 1

print(f"Process completed.")
print(f"Photos Deleted: {photos_deleted}")
print(f"Photos Remaining: {photos_remaining}")

If I see images with the Sample Tag or with more than 900,000 KB, what I do is restore them then manually resize them by replacing the lowest number with 1024. (or see out the original sized images for the Samples).

If somehow I end up with more than half my images deleted, even with this, what I would then do is bring the deleted images to a new folder and run this code which will automatically resize them by a certain percentage.

# Define the path to your image folder on Google Drive (replace with your actual path)
image_folder_path = '/content/drive/MyDrive/Loras/[INSERT FOLDER HERE]/ToResize'

# Define the resize factor (The number after 1. would be your percentage, with 2 being outright doubling and so forth.)
resize_factor = 1.25

# Loop through all files in the image folder
for filename in os.listdir(image_folder_path):
  # Check if it's an image file
  if filename.lower().endswith(('.jpg', '.jpeg', '.png')):
    # Open the image
    img = Image.open(os.path.join(image_folder_path, filename))

    # Get image size
    width, height = img.size

    # Resize the image
    new_width = int(width * resize_factor)
    new_height = int(height * resize_factor)
    resized_img = img.resize((new_width, new_height), resample=Image.LANCZOS)  # Use LANCZOS resampling

    # Save the resized image with a new name (optional)
    # new_filename = f"{filename[:-4]}_resized.{filename[-3:]}"
    new_filename = filename  # Save over the original file (comment out the line above)

    # Save the resized image
    resized_img.save(os.path.join(image_folder_path, new_filename))

print("Images resized successfully!")

By how much is up to you, but I prefer 25% if it's only a few more images you need and up to 2 times the original size if you want as many images as you can. Anything after 2 times would be overkill, especially since the next step I do is to bring them back into the dataset where I run the deletion again, this time observing which ones still go and repeating the process.

Or, if I were to put this in BASIC terms.

10 REM Start Deletion Process
20 GOSUB DeleteCode
30 IF SampleImageFound THEN GOSUB AddToDataset
40 IF ImageSize > 900000 THEN GOSUB RestoreAndResize
50 IF ImagesRemaining < TotalImages / 2 THEN GOSUB RestoreAndAdd
60 GOSUB Tagging
70 END

80 DeleteCode:
90 REM Deletion code logic here
100 RETURN

110 AddToDataset:
120 REM Logic to find original image and add to dataset
130 RETURN

140 RestoreAndResize:
150 REM Logic to restore and resize image
160 RETURN

170 RestoreAndAdd:
180 REM Logic to restore images and add to restore images
190 REM Call resize code
200 GOSUB ResizeCode
210 RETURN

220 ResizeCode:
230 REM Logic to resize image
240 RETURN

250 Tagging:
260 REM Tagging code logic here
270 RETURN