The overall workflow of the training and inference process are provided in Alg. 1 and Alg. 2. Model Architecture Following [ 9 ], we use a U-Net with 4-channel input and 1-channel output. Both input and output resolution is set to 256 256 . Training Settings All experiments are conducted on 8 NVIDIA RTX3090 GPUs with Pytorch. After a complete reverse diffusion process, the output is resized to the original size. We apply Non-Maximum Suppression (NMS, with 0.3 as threshold) to these patches to remove Our SegRefiner can robustly correct prediction errors both outside and inside the coarse mask.

While state-of-the-art background removal models excel at realistic imagery, they frequently underperform in specialized domains such as anime-style content, where complex features like hair and transparency present unique challenges. To address this limitation, we collected and annotated a custom dataset of 1,228 high-quality anime images of characters and objects, and fine-tuned the open-sourced BiRefNet model on this dataset. This resulted in marked improvements in background removal accuracy for anime-style images, increasing from 95.3% to 99.5% for our newly introduced Pixel Accuracy metric. We are open-sourcing the code, the fine-tuned model weights, as well as the dataset at: https://github.com/MatteoKartoon/BiRefNet.

Dichotomous Image Segmentation (DIS) tasks require highly precise annotations, and traditional dataset creation methods are labor intensive, costly, and require extensive domain expertise. Although using synthetic data for DIS is a promising solution to these challenges, current generative models and techniques struggle with the issues of scene deviations, noise-induced errors, and limited training sample variability. To address these issues, we introduce a novel approach, Mask Factory, which provides a scalable solution for generating diverse and precise datasets, markedly reducing preparation time and costs. We first introduce a general mask editing method that combines rigid and non-rigid editing techniques to generate high-quality synthetic masks. Specially, rigid editing leverages geometric priors from diffusion models to achieve precise viewpoint transformations under zero-shot conditions, while non-rigid editing employs adversarial training and self-attention mechanisms for complex, topologically consistent modifications.