fix generation and further cleanup by moving generate method into Dal…

make sure generation is done in eval mode, and sets it back to origin…

use state of the art vision transformer in CLIP - 0.0.18

0.0.17

0.0.2

release development version where vae and decoder is trained end-to-end

make sure last token can still predict eos

allow for transparent images

0.0.14

use logits masking to ensure image is produced after text