Closed enmaru0 closed 1 week ago
Hi @enmaru0 Thank you for your interest in our work. We apologize for the delayed response; we have been fully occupied with research tasks. Although we have not yet investigated our method for super-resolution tasks, you can use our model for this purpose. Ensure you have a dataset containing high-quality images, as this will help your model converge quickly. If your dataset is of poor quality and contains artifacts and noise, the model may struggle to converge. Additionally, since you are changing the model input mask to a low-resolution image, it might be beneficial to include some simple augmentations like flipping and rotating during training. We are currently working on the next version of our method to improve its ability to converge with noisy and poor-quality datasets.
Thanks for sharing your great work. You mentioned that this work was inspired by a super-resolution technique in the paper. I am reversely trying to modify the code to train 3D super resolution task. Should I just replace the input mask (2 channels) to a low-resolution image which is up-sampled (1 channels) and concatenated to the noise-image to train super resolution model? Or do I need to modify any other parts?
If possible, I hope you can provide some advice on this matter. Thank you so much for your kindest help.