Dear author,
May I ask if the prompt editing function for Gaussian splatting is achieved by directly computing the similarity between the language feature of each Gaussian and the prompt feature? If that is the case, during training, only the alignment between the rendered feature map and the text features is supervised, without specifically aligning the language features of individual Gaussians. Given this, can effective alignment still be directly achieved during inference?
Thanks!
Dear author,
May I ask if the prompt editing function for Gaussian splatting is achieved by directly computing the similarity between the language feature of each Gaussian and the prompt feature? If that is the case, during training, only the alignment between the rendered feature map and the text features is supervised, without specifically aligning the language features of individual Gaussians. Given this, can effective alignment still be directly achieved during inference?
Thanks!