I was wondering if one can take an image generated by some model, match it with some other reference image, compute a loss on the matches' pixel distance, and propagate gradient updates to the base model that generated the image.
My use case is in differentiable rendering, where I generate an image from a mesh, and want to update the mesh's location.
I was wondering if one can take an image generated by some model, match it with some other reference image, compute a loss on the matches' pixel distance, and propagate gradient updates to the base model that generated the image.
My use case is in differentiable rendering, where I generate an image from a mesh, and want to update the mesh's location.
Thanks.