Imagebucket
Upload Images
Upload , Crop, Save All
What is Imagebucket?
Preparing images for AI training can be a time-consuming task. Most existing cropping tools are designed for image editing or for processing large datasets for optimization purposes. I created Imagebucket to address this by enabling users to assemble datasets quickly and consistently.
Tips for Training a LoRA/Model:
A balanced dataset is key. For your LoRA type, pick images that consistently show your style, character, or concept, but mix up other details. If you have enough images, aim for even variety—like full-body shots, portraits, and close-ups for a character, covering different angles. Unvaried elements might sneak into the training. With fewer images, focus on keeping only the target style, character, or concept as the common thread.
Privacy:
Imagebucket leverages cropper.js to adjust and trim your images directly in your browser. Everything runs locally on your device, so no photos are sent or upload to a server.
Do you need a 1:1 square for training a LoRA or checkpoint?
No, with “Enable buckets” on, you can use different aspect ratios. Aim for at least two images per bucket—preferably an even number—and balance them if you have extra images. Don’t overthink it; imperfect bucketing won’t ruin your LoRA, though it helps efficiency. If stuck, crop to 1:1 squares for simplicity and max pixels.
What if I don’t have a lot of data to train?:
When training a LoRA with a small number of images, the key is to maximize what you have through strategic cropping and bucketing. For a Style LoRA, using a single high-resolution image like Caravaggio's "The Supper at Emmaus," you can create multiple training samples by focusing on the stylistic elements that define Caravaggio’s work—think dramatic lighting, chiaroscuro, and realistic details. For Character LoRAs, take a full-body shot and crop to cowboy shots, headshots, or close-ups of iconic details like clothing or accessories (e.g., weapons or pendants) to teach the model those specifics.
Image quality:
JPG quality is set to 100% to ensure maximum image fidelity during training. PNG quality is default, with no quality lost, as it remains lossless.
version v0.04
Donation:
ETH:
BTC:
