Ablating Instances (click images to see more examples)
We show qualitative comparison of our model-based concept ablation method (2nd row), with our noise-based variant (3rd row) and the baseline method of maximizing the loss (4rth row) on ablating various instances. In our noise-based variant we fine-tune the model on the redefined training pair of (target concept caption, anchor concept image). On target concept the more different the samples compared to one by pretrained model the better. On surrounding concept the more similar the samples to the one generated by pretrained model the better.
