How to use Imagebucket:

Upload:

Hit the "Upload" button to add multiple images from your device.

  • upload_01
  • upload_04
Crop:

Once your images are uploaded, use the cropper tool to select and frame the specific area of each image you want the model to focus on. This helps highlight the most relevant parts of your images for training.

  • cropping_01
Aspect Ratio and Bucketing:

By default, Bucket feature is on by default in platforms like Kohya_SS and Civitai’s online training. This feature will allow training for images that are in different aspect ratios and resolution. Imagebucket allows users to pick different aspect ratios for each individual image.

  • bucket
Duplicating Images:

The image management tool includes a "Duplicate" button, allowing you to quickly create copies of your existing images to expand your dataset. This is especially helpful if you’re working with a limited number of unique images.

  • duplicate

By cropping different areas or applying different aspect ratios, you can create variations that emphasize specific details. While this approach can help the model learn particular features, it’s generally better to have a large and diverse dataset of unique images for improved generalization and performance in your trained model.

  • duplicrop

Example:

N.C. Wyeth King Arthu Lora Model

I use N.C Wyeth's painting for this Lora training example. This is a lora model train on Pony, with unique 12 images, and 32 total training images after cropping.

V1 Data set:
Cropped Image 1 Cropped Image 2 Cropped Image 3 Cropped Image 4 Cropped Image 5 Cropped Image 6 Cropped Image 7 Cropped Image 8 Cropped Image 9 Cropped Image 10 Cropped Image 11 Cropped Image 12 Cropped Image 13 Cropped Image 14 Cropped Image 15 Cropped Image 16 Cropped Image 17 Cropped Image 18 Cropped Image 19 Cropped Image 20 Cropped Image 21 Cropped Image 22 Cropped Image 23 Cropped Image 24 Cropped Image 25 Cropped Image 26 Cropped Image 27 Cropped Image 28 Cropped Image 29 Cropped Image 30 Cropped Image 31 Cropped Image 32
Large Image
V1 Result:
Image 1 Thumbnail Image 2 Thumbnail Image 3 Thumbnail
Large Image

Generated in Pony

Image 1 prompt: a knight holding a shield facing a red sky with a black sun, helmet, holding, sword, shiled, weapon, holding weapon, red sky, black sun, painting, traditional media, high contrast, best quality

Image 2 prompt: 1boy, masterpiece, best quality, good quality, knight, newest, highres, absurdres, holding, weapon, boots, sword, white dress, red armor, helmet, gauntlets, shield, sun, blue sky, full armor

Image 3 prompt: a knight holding a shield in dark green fancy armor, black sky, high quality, fantasy, dramatic lighting, high res, darkness, horse, armor

N.C. Wyeth King Arthu V1 Lora/Data Download Link:

Lora and Training Data Download link

V1 Training Afterthoughts:

The limited dataset seems to perform remarkably well in capturing N.C. Wyeth's style, even within a model like Pony. Adding more images could enhance flexibility, but with just 12 images, I think the results are incredible.

version v0.04