VQGAN model sporadically produces empty results

#13
by dwieker - opened

I'm generating image encodings with dalle-mini/dalle-mini/mega-1-fp16:latest and decoding with dalle-mini/vqgan_imagenet_f16_16384. About 70% of the time VQGAN generates a blank image:

[[0. 0. 0.]
[0. 0. 0.]
[0. 0. 0.]
...
[0. 0. 0.]
[0. 0. 0.]
[0. 0. 0.]]

[[0. 0. 0.]
[0. 0. 0.]
[0. 0. 0.]
...
[0. 0. 0.]
[0. 0. 0.]
[0. 0. 0.]]

[[0. 0. 0.]
[0. 0. 0.]
[0. 0. 0.]
...
[0. 0. 0.]
[0. 0. 0.]
[0. 0. 0.]]]]

but other times is successful -- on the same exact prompt.

dwieker changed discussion status to closed
DALLE mini org

Make sure you use full precision for the VQGAN and it should not happen

Is not working

Sign up or log in to comment