So much for “generation” - it seems as if these models are just overfitting on extremely small subset of input data that it did not utterly failed to train on, almost that there could be geniuses who would be able to directly generate weight data from said images without all the gradient descent thing.