We refer to depthfm to perform the portrait matting task. We utilize the generalization and robustness capabilities obtained after training with stable diffusion on a large amount of data. We use the principle of flow matching to fine-tune the pre-trained model with less data. Summary: Achieve better robustness with less data and less training time.
我们参考depthfm来做人像matting任务，利用stable diffusion在大量数据上训练后获得的泛化，鲁棒性能力，使用flow matching原理来使用更少数据微调预训练的模型。总结：用更少的数据，更少的训练时间，达到更好的鲁棒性
Please refer to the depthfm project and install the dependent software.
download the pretrained model from BaiDu passwd：6s57, and put in exp/matting.
python inference_matting.py \
--num_steps 2 \
--ensemble_size 4 \
--ckpt ${MODEL_PATH}the matting results
We refer to the papers LFM and depthfm and perform simple fine-tuning on the P3M10K dataset. For specific details, please modify the code in the train_matting.py script.
我们参看LFM和depthfm论文在P3M10K数据集上进行简单微调.具体细节请修改train_matting.py脚本中的代码。
the code main from depthfm and LFM, Thank them for their excellent work.