SoloCropper is a high-efficiency human image cropping tool tailored for dataset preparation. It automates the generation of multi-spec outputs from multiple images, allowing users to crop various body segments and sizes in a single pass based on custom parameters.
I know there are already many image cropping tools out there. As a LoRA maker, cropping images is an unavoidable part of the workflow. I’ve tried numerous automated tools in the past, but I always found myself returning to manual cropping because it was the only way to get the exact results I wanted.
I put together SoloCropper to bridge that gap. It can rapidly generate a large volume of candidate crops, many of which now meet my personal standards for dataset quality. While it doesn't completely replace manual cropping for those tricky shots, it has significantly reduced my overall workload. Since it has successfully met the requirements of my own workflow, I believe it can offer some assistance to others building their own datasets.
Here is a brief overview of what SoloCropper can do:
## I. Fast & Accurate Detection
Powered by high-efficiency Ultralytics YOLO models, SoloCropper delivers rapid human recognition even on standard CPUs.
The models I use are yolo26x-seg.pt and yolo26x-pose.pt. At just over 200MB, they are lightweight enough to run remarkably fast on my entry-level laptop, providing high accuracy with very few errors.
II. A Wide Range of Output Choices
You can easily customize numerous output settings via the configuration file, including crop positions, aspect ratios, output dimensions, image formats, and compression levels. You can even specify multiple sizes or ratios for each crop location.
You can also choose to output wireframe previews for inspection instead of actually cropping the images.
## III. 8 Box Group Definitions
SoloCropper currently provides 8 different cropping boxes for various body parts, including:
- full: Full-body box.
- shoulder: From the top of the person down to the shoulder line.
- shoulder_only: A narrower shoulder-focused region that uses the shoulder points as its boundary and does not widen its horizontal range because of arms, clothing, or hair.
- waist: From the top of the person down to the waist line (estimated, not always accurate).
- hip: From the top of the person down to the hip line.
- glute: From the top of the person down to the glute line (estimated, not always accurate).
- knee: From the top of the person down to the knee line.
- ankle: From the top of the person down to the ankle line.
Except for waist and glute, which are estimated values, the other six boxes are based on precise locations identified by the YOLO model.
IV. Limitations
Due to the detection limits of the YOLO models, SoloCropper performs best on real-life photos. It also handles 3D characters and realistic illustrations well. However, its performance on 2D anime or anthropomorphic animals is less consistent, it may only detect the overall silhouette without being able to distinguish specific body parts.
Currently, SoloCropper is limited to human image recognition and cropping. I don't have a clear plan for non-human images just yet, but if a good idea comes up, I will definitely expand the tool’s capabilities.
SoloCropper is available on GitHub for anyone interested: https://github.com/sololo-xyz/SoloCropper
Detailed configuration instructions are available in Config-Guide-EN.md: https://github.com/sololo-xyz/SoloCropper/blob/main/docs/Config-Guide-EN.md
