Tools to Save You Time Preparing Datasets
I've been developing Windows software for nearly 20 years, mostly with one goal: eliminating repetitive, time-consuming work. Asides for breathing, nothing seems more repetitive than the find, sort, curate, crop, caption, refine, train, test, rinse, and repeat of building training datasets.
Since diving hard into generative AI training in late 2022, most of what I've built has focused on image dataset management, organization, manipulation, and captioning. Many of these started as tools to speed up LoRA and finetune creation, but grew in feature count once I realized how much time they were saving me.
Here's a quick overview of what's currently available. All of my tools are portable, with no install required. Also, my tools are 100% offline except the couple that perform a one-time download of needed model(s) from HuggingFace when they run the first time.
Caption Forge - Image Search Engine, Organizer & Tagger
CaptionForge indexes your image collection and lets you search it using tags or plain-language descriptions without moving, copying, or modifying your images.
Instead of manually digging through thousands of images to build a dataset, you can search for what you need and export the matching images, and optionally with their .txt caption files, with a single click.
Caption Cleaner - Bulk Caption Editor & Dataset Organizer
CaptionCleaner lets you manage an entire dataset's captions and tags from one interface instead of opening and editing thousands of individual text files.
You can bulk rename tags, fix typos, remove unwanted tags, and organize images, making dataset cleanup considerably faster.
AutoCrop Anything - Bulk Object Detection & Cropping
AutoCrop Anything uses YOLO/Ultralytics models to detect and crop virtually any trained object, person, or body part in bulk.
It can process large numbers of images automatically, organize the cropped output, preserve the original image dimensions when desired, and supports compatible YOLO/Ultralytics '.pt' models already used by tools such as aDetailer.
Ultimate FaceCropper - Bulk Face Detection, Cropping & Masking
Ultimate FaceCropper automatically detects and crops faces in bulk, with optional masking, resizing, face recovery, multi-face support, and automatic output organization.
It also supports CUDA acceleration, configurable batch size and thread count, and masks for multiple detected image elements, removing much of the manual work involved in preparing face datasets.
Bucket Commander - Image Bucketing & Bulk Cropping
My personal favorite. Bucket Commander automatically organizes images by orientation, aspect ratio, or dimensions, making it easier to prepare datasets for trainers such as Kohya, OneTrainer, and AI-Toolkit.
It can also bulk-crop images to specific aspect ratios from multiple starting points, or generate every available crop automatically, so you can choose the crops that work best for you.
Stable Diffusion v1.5 LoRA Training Calculator
A little late for SD1.5, but this calculator is something I've spent the last three years refining through research and comparative testing.
It helps reduce the guesswork involved in planning SD1.5 LoRA training by accounting for factors such as dataset size, concept complexity, multiple concepts, and regularization to produce viable training recommendations for scheduler, learning rate cycles, warm up steps, U-Net LR, TE LR, epochs, batch size, repeats, etc.
Depending on how this lands with users, I plan to add LoRA and finetune calculations for SDXL, Flux, and Krea in the future.
Additionally, if you have suggestions for new features that you'd like to see added to any of my tools, let me know. I will certainly consider them.