For many everyday images, background removal works well. The remaining challenge is images with fine hair, fur, glass, transparent objects, motion blur, low-contrast subjects, or cluttered backgrounds. Improving results on those difficult cases has been the focus of this project.This is the third iteration of the model, built through many rounds of training and evaluation. I ve focused on improving performance on the kinds of images that are still difficult for many background removal tools.I recently released the third iteration of my open-weight background removal model (Apache-2.0 for my contributions; see repository for third-party licensing). My goal is to make high-quality background removal easy to run locally (or on your own infra)Some example results: - Open model results: https://withoutbg.com/open-model/results - Comparison with remove.bg: https://withoutbg.com/compare/withoutbg-open-model-vs-remove...(The comparison intentionally focuses on challenging images, since many tools produce similar results on easier ones. I was more interested in showing where the approaches differ)You can run the model in several ways: - macOS app: https://withoutbg.com/mac - GIMP plugin: https://github.com/withoutbg/withoutbg-gimp - Python package: https://github.com/withoutbg/withoutbg-python - Docker image/service: https://withoutbg.com/docs/open-model/docker - Model weights (Hugging Face): https://huggingface.co/withoutbg/withoutbg-openweights-onnxI would be interested in feedback on both successes and failures. If you have images where the model produces poor results (or unexpectedly good ones compared to other tools) I would appreciate seeing/knowing them. Those kinds of examples have been the most useful for improving the model. I will be around to answer questions about the model, implementation, and the decisions behind it.