Hi, I see that most pre-trained unconditional LDMs are trained on just one type of image such as “flowers”, “landscape”,… Therefore, they can only generate images related to one specific class (e.g., generating different flowers, but still flowers, not a car or a dog).
If we train an unconditional LDM on images of different datasets that include various classes, is it ok? Depending on the random seed, maybe sometimes it would generate a flower (in a forest), but sometimes a dog (sitting in a car)?
1 Like