Closed ninjasaid2k closed 3 months ago
Theoretically, our proposed robust unlearning framework could potentially be applied to both SD v2.0 and SDXL. However, the current codebase supports only SD v1.x, which is the focus of existing DM unlearning methods. For SD v1.x, fine-tuning through AdvUnlearn with a batch size of 5 for utility-retaining regularization, requires approximately 40GB.
Can this work for different model types like 2.0 and xl? and what the VRAM requirement?