Performance compared to dalle-mini

[kargarisaac]

Hello,

I'm trying to compare the performance of dalle2 (this repo - colab) with dalle-mini (https://huggingface.co/spaces/dalle-mini/dalle-mini). The dalle-mini performs much better. I cannot get meaningful images from dalle2. Is there an specific setting I have to do to get a good result? Or it is not completely trained? Or sth else?

for example: "a programmer sitting in a coffeeshop."

dalle2 output:

dalle-mini output:

[nousr]

Is there an specific setting I have to do to get a good result? Or it is not completely trained? Or sth else?

We are still working towards a fully trained model--tools and supporting code are being worked on to slingshot our dalle2 recreation into a much more usable state, everything right now is still very early on in development (relatively speaking) compared to dalle-mini!

Stay tuned :D